1 /* 2 ** 2001 September 15 3 ** 4 ** The author disclaims copyright to this source code. In place of 5 ** a legal notice, here is a blessing: 6 ** 7 ** May you do good and not evil. 8 ** May you find forgiveness for yourself and forgive others. 9 ** May you share freely, never taking more than you give. 10 ** 11 ************************************************************************* 12 ** This module contains C code that generates VDBE code used to process 13 ** the WHERE clause of SQL statements. This module is responsible for 14 ** generating the code that loops through a table looking for applicable 15 ** rows. Indices are selected and used to speed the search when doing 16 ** so is applicable. Because this module is responsible for selecting 17 ** indices, you might also think of this module as the "query optimizer". 18 */ 19 #include "sqliteInt.h" 20 #include "whereInt.h" 21 22 /* 23 ** Extra information appended to the end of sqlite3_index_info but not 24 ** visible to the xBestIndex function, at least not directly. The 25 ** sqlite3_vtab_collation() interface knows how to reach it, however. 26 ** 27 ** This object is not an API and can be changed from one release to the 28 ** next. As long as allocateIndexInfo() and sqlite3_vtab_collation() 29 ** agree on the structure, all will be well. 30 */ 31 typedef struct HiddenIndexInfo HiddenIndexInfo; 32 struct HiddenIndexInfo { 33 WhereClause *pWC; /* The Where clause being analyzed */ 34 Parse *pParse; /* The parsing context */ 35 }; 36 37 /* Forward declaration of methods */ 38 static int whereLoopResize(sqlite3*, WhereLoop*, int); 39 40 /* 41 ** Return the estimated number of output rows from a WHERE clause 42 */ 43 LogEst sqlite3WhereOutputRowCount(WhereInfo *pWInfo){ 44 return pWInfo->nRowOut; 45 } 46 47 /* 48 ** Return one of the WHERE_DISTINCT_xxxxx values to indicate how this 49 ** WHERE clause returns outputs for DISTINCT processing. 50 */ 51 int sqlite3WhereIsDistinct(WhereInfo *pWInfo){ 52 return pWInfo->eDistinct; 53 } 54 55 /* 56 ** Return the number of ORDER BY terms that are satisfied by the 57 ** WHERE clause. A return of 0 means that the output must be 58 ** completely sorted. A return equal to the number of ORDER BY 59 ** terms means that no sorting is needed at all. A return that 60 ** is positive but less than the number of ORDER BY terms means that 61 ** block sorting is required. 62 */ 63 int sqlite3WhereIsOrdered(WhereInfo *pWInfo){ 64 return pWInfo->nOBSat; 65 } 66 67 /* 68 ** In the ORDER BY LIMIT optimization, if the inner-most loop is known 69 ** to emit rows in increasing order, and if the last row emitted by the 70 ** inner-most loop did not fit within the sorter, then we can skip all 71 ** subsequent rows for the current iteration of the inner loop (because they 72 ** will not fit in the sorter either) and continue with the second inner 73 ** loop - the loop immediately outside the inner-most. 74 ** 75 ** When a row does not fit in the sorter (because the sorter already 76 ** holds LIMIT+OFFSET rows that are smaller), then a jump is made to the 77 ** label returned by this function. 78 ** 79 ** If the ORDER BY LIMIT optimization applies, the jump destination should 80 ** be the continuation for the second-inner-most loop. If the ORDER BY 81 ** LIMIT optimization does not apply, then the jump destination should 82 ** be the continuation for the inner-most loop. 83 ** 84 ** It is always safe for this routine to return the continuation of the 85 ** inner-most loop, in the sense that a correct answer will result. 86 ** Returning the continuation the second inner loop is an optimization 87 ** that might make the code run a little faster, but should not change 88 ** the final answer. 89 */ 90 int sqlite3WhereOrderByLimitOptLabel(WhereInfo *pWInfo){ 91 WhereLevel *pInner; 92 if( !pWInfo->bOrderedInnerLoop ){ 93 /* The ORDER BY LIMIT optimization does not apply. Jump to the 94 ** continuation of the inner-most loop. */ 95 return pWInfo->iContinue; 96 } 97 pInner = &pWInfo->a[pWInfo->nLevel-1]; 98 assert( pInner->addrNxt!=0 ); 99 return pInner->addrNxt; 100 } 101 102 /* 103 ** While generating code for the min/max optimization, after handling 104 ** the aggregate-step call to min() or max(), check to see if any 105 ** additional looping is required. If the output order is such that 106 ** we are certain that the correct answer has already been found, then 107 ** code an OP_Goto to by pass subsequent processing. 108 ** 109 ** Any extra OP_Goto that is coded here is an optimization. The 110 ** correct answer should be obtained regardless. This OP_Goto just 111 ** makes the answer appear faster. 112 */ 113 void sqlite3WhereMinMaxOptEarlyOut(Vdbe *v, WhereInfo *pWInfo){ 114 WhereLevel *pInner; 115 int i; 116 if( !pWInfo->bOrderedInnerLoop ) return; 117 if( pWInfo->nOBSat==0 ) return; 118 for(i=pWInfo->nLevel-1; i>=0; i--){ 119 pInner = &pWInfo->a[i]; 120 if( (pInner->pWLoop->wsFlags & WHERE_COLUMN_IN)!=0 ){ 121 sqlite3VdbeGoto(v, pInner->addrNxt); 122 return; 123 } 124 } 125 sqlite3VdbeGoto(v, pWInfo->iBreak); 126 } 127 128 /* 129 ** Return the VDBE address or label to jump to in order to continue 130 ** immediately with the next row of a WHERE clause. 131 */ 132 int sqlite3WhereContinueLabel(WhereInfo *pWInfo){ 133 assert( pWInfo->iContinue!=0 ); 134 return pWInfo->iContinue; 135 } 136 137 /* 138 ** Return the VDBE address or label to jump to in order to break 139 ** out of a WHERE loop. 140 */ 141 int sqlite3WhereBreakLabel(WhereInfo *pWInfo){ 142 return pWInfo->iBreak; 143 } 144 145 /* 146 ** Return ONEPASS_OFF (0) if an UPDATE or DELETE statement is unable to 147 ** operate directly on the rowids returned by a WHERE clause. Return 148 ** ONEPASS_SINGLE (1) if the statement can operation directly because only 149 ** a single row is to be changed. Return ONEPASS_MULTI (2) if the one-pass 150 ** optimization can be used on multiple 151 ** 152 ** If the ONEPASS optimization is used (if this routine returns true) 153 ** then also write the indices of open cursors used by ONEPASS 154 ** into aiCur[0] and aiCur[1]. iaCur[0] gets the cursor of the data 155 ** table and iaCur[1] gets the cursor used by an auxiliary index. 156 ** Either value may be -1, indicating that cursor is not used. 157 ** Any cursors returned will have been opened for writing. 158 ** 159 ** aiCur[0] and aiCur[1] both get -1 if the where-clause logic is 160 ** unable to use the ONEPASS optimization. 161 */ 162 int sqlite3WhereOkOnePass(WhereInfo *pWInfo, int *aiCur){ 163 memcpy(aiCur, pWInfo->aiCurOnePass, sizeof(int)*2); 164 #ifdef WHERETRACE_ENABLED 165 if( sqlite3WhereTrace && pWInfo->eOnePass!=ONEPASS_OFF ){ 166 sqlite3DebugPrintf("%s cursors: %d %d\n", 167 pWInfo->eOnePass==ONEPASS_SINGLE ? "ONEPASS_SINGLE" : "ONEPASS_MULTI", 168 aiCur[0], aiCur[1]); 169 } 170 #endif 171 return pWInfo->eOnePass; 172 } 173 174 /* 175 ** Return TRUE if the WHERE loop uses the OP_DeferredSeek opcode to move 176 ** the data cursor to the row selected by the index cursor. 177 */ 178 int sqlite3WhereUsesDeferredSeek(WhereInfo *pWInfo){ 179 return pWInfo->bDeferredSeek; 180 } 181 182 /* 183 ** Move the content of pSrc into pDest 184 */ 185 static void whereOrMove(WhereOrSet *pDest, WhereOrSet *pSrc){ 186 pDest->n = pSrc->n; 187 memcpy(pDest->a, pSrc->a, pDest->n*sizeof(pDest->a[0])); 188 } 189 190 /* 191 ** Try to insert a new prerequisite/cost entry into the WhereOrSet pSet. 192 ** 193 ** The new entry might overwrite an existing entry, or it might be 194 ** appended, or it might be discarded. Do whatever is the right thing 195 ** so that pSet keeps the N_OR_COST best entries seen so far. 196 */ 197 static int whereOrInsert( 198 WhereOrSet *pSet, /* The WhereOrSet to be updated */ 199 Bitmask prereq, /* Prerequisites of the new entry */ 200 LogEst rRun, /* Run-cost of the new entry */ 201 LogEst nOut /* Number of outputs for the new entry */ 202 ){ 203 u16 i; 204 WhereOrCost *p; 205 for(i=pSet->n, p=pSet->a; i>0; i--, p++){ 206 if( rRun<=p->rRun && (prereq & p->prereq)==prereq ){ 207 goto whereOrInsert_done; 208 } 209 if( p->rRun<=rRun && (p->prereq & prereq)==p->prereq ){ 210 return 0; 211 } 212 } 213 if( pSet->n<N_OR_COST ){ 214 p = &pSet->a[pSet->n++]; 215 p->nOut = nOut; 216 }else{ 217 p = pSet->a; 218 for(i=1; i<pSet->n; i++){ 219 if( p->rRun>pSet->a[i].rRun ) p = pSet->a + i; 220 } 221 if( p->rRun<=rRun ) return 0; 222 } 223 whereOrInsert_done: 224 p->prereq = prereq; 225 p->rRun = rRun; 226 if( p->nOut>nOut ) p->nOut = nOut; 227 return 1; 228 } 229 230 /* 231 ** Return the bitmask for the given cursor number. Return 0 if 232 ** iCursor is not in the set. 233 */ 234 Bitmask sqlite3WhereGetMask(WhereMaskSet *pMaskSet, int iCursor){ 235 int i; 236 assert( pMaskSet->n<=(int)sizeof(Bitmask)*8 ); 237 for(i=0; i<pMaskSet->n; i++){ 238 if( pMaskSet->ix[i]==iCursor ){ 239 return MASKBIT(i); 240 } 241 } 242 return 0; 243 } 244 245 /* 246 ** Create a new mask for cursor iCursor. 247 ** 248 ** There is one cursor per table in the FROM clause. The number of 249 ** tables in the FROM clause is limited by a test early in the 250 ** sqlite3WhereBegin() routine. So we know that the pMaskSet->ix[] 251 ** array will never overflow. 252 */ 253 static void createMask(WhereMaskSet *pMaskSet, int iCursor){ 254 assert( pMaskSet->n < ArraySize(pMaskSet->ix) ); 255 pMaskSet->ix[pMaskSet->n++] = iCursor; 256 } 257 258 /* 259 ** If the right-hand branch of the expression is a TK_COLUMN, then return 260 ** a pointer to the right-hand branch. Otherwise, return NULL. 261 */ 262 static Expr *whereRightSubexprIsColumn(Expr *p){ 263 p = sqlite3ExprSkipCollateAndLikely(p->pRight); 264 if( ALWAYS(p!=0) && p->op==TK_COLUMN && !ExprHasProperty(p, EP_FixedCol) ){ 265 return p; 266 } 267 return 0; 268 } 269 270 /* 271 ** Advance to the next WhereTerm that matches according to the criteria 272 ** established when the pScan object was initialized by whereScanInit(). 273 ** Return NULL if there are no more matching WhereTerms. 274 */ 275 static WhereTerm *whereScanNext(WhereScan *pScan){ 276 int iCur; /* The cursor on the LHS of the term */ 277 i16 iColumn; /* The column on the LHS of the term. -1 for IPK */ 278 Expr *pX; /* An expression being tested */ 279 WhereClause *pWC; /* Shorthand for pScan->pWC */ 280 WhereTerm *pTerm; /* The term being tested */ 281 int k = pScan->k; /* Where to start scanning */ 282 283 assert( pScan->iEquiv<=pScan->nEquiv ); 284 pWC = pScan->pWC; 285 while(1){ 286 iColumn = pScan->aiColumn[pScan->iEquiv-1]; 287 iCur = pScan->aiCur[pScan->iEquiv-1]; 288 assert( pWC!=0 ); 289 do{ 290 for(pTerm=pWC->a+k; k<pWC->nTerm; k++, pTerm++){ 291 if( pTerm->leftCursor==iCur 292 && pTerm->u.x.leftColumn==iColumn 293 && (iColumn!=XN_EXPR 294 || sqlite3ExprCompareSkip(pTerm->pExpr->pLeft, 295 pScan->pIdxExpr,iCur)==0) 296 && (pScan->iEquiv<=1 || !ExprHasProperty(pTerm->pExpr, EP_FromJoin)) 297 ){ 298 if( (pTerm->eOperator & WO_EQUIV)!=0 299 && pScan->nEquiv<ArraySize(pScan->aiCur) 300 && (pX = whereRightSubexprIsColumn(pTerm->pExpr))!=0 301 ){ 302 int j; 303 for(j=0; j<pScan->nEquiv; j++){ 304 if( pScan->aiCur[j]==pX->iTable 305 && pScan->aiColumn[j]==pX->iColumn ){ 306 break; 307 } 308 } 309 if( j==pScan->nEquiv ){ 310 pScan->aiCur[j] = pX->iTable; 311 pScan->aiColumn[j] = pX->iColumn; 312 pScan->nEquiv++; 313 } 314 } 315 if( (pTerm->eOperator & pScan->opMask)!=0 ){ 316 /* Verify the affinity and collating sequence match */ 317 if( pScan->zCollName && (pTerm->eOperator & WO_ISNULL)==0 ){ 318 CollSeq *pColl; 319 Parse *pParse = pWC->pWInfo->pParse; 320 pX = pTerm->pExpr; 321 if( !sqlite3IndexAffinityOk(pX, pScan->idxaff) ){ 322 continue; 323 } 324 assert(pX->pLeft); 325 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 326 if( pColl==0 ) pColl = pParse->db->pDfltColl; 327 if( sqlite3StrICmp(pColl->zName, pScan->zCollName) ){ 328 continue; 329 } 330 } 331 if( (pTerm->eOperator & (WO_EQ|WO_IS))!=0 332 && (pX = pTerm->pExpr->pRight)->op==TK_COLUMN 333 && pX->iTable==pScan->aiCur[0] 334 && pX->iColumn==pScan->aiColumn[0] 335 ){ 336 testcase( pTerm->eOperator & WO_IS ); 337 continue; 338 } 339 pScan->pWC = pWC; 340 pScan->k = k+1; 341 return pTerm; 342 } 343 } 344 } 345 pWC = pWC->pOuter; 346 k = 0; 347 }while( pWC!=0 ); 348 if( pScan->iEquiv>=pScan->nEquiv ) break; 349 pWC = pScan->pOrigWC; 350 k = 0; 351 pScan->iEquiv++; 352 } 353 return 0; 354 } 355 356 /* 357 ** This is whereScanInit() for the case of an index on an expression. 358 ** It is factored out into a separate tail-recursion subroutine so that 359 ** the normal whereScanInit() routine, which is a high-runner, does not 360 ** need to push registers onto the stack as part of its prologue. 361 */ 362 static SQLITE_NOINLINE WhereTerm *whereScanInitIndexExpr(WhereScan *pScan){ 363 pScan->idxaff = sqlite3ExprAffinity(pScan->pIdxExpr); 364 return whereScanNext(pScan); 365 } 366 367 /* 368 ** Initialize a WHERE clause scanner object. Return a pointer to the 369 ** first match. Return NULL if there are no matches. 370 ** 371 ** The scanner will be searching the WHERE clause pWC. It will look 372 ** for terms of the form "X <op> <expr>" where X is column iColumn of table 373 ** iCur. Or if pIdx!=0 then X is column iColumn of index pIdx. pIdx 374 ** must be one of the indexes of table iCur. 375 ** 376 ** The <op> must be one of the operators described by opMask. 377 ** 378 ** If the search is for X and the WHERE clause contains terms of the 379 ** form X=Y then this routine might also return terms of the form 380 ** "Y <op> <expr>". The number of levels of transitivity is limited, 381 ** but is enough to handle most commonly occurring SQL statements. 382 ** 383 ** If X is not the INTEGER PRIMARY KEY then X must be compatible with 384 ** index pIdx. 385 */ 386 static WhereTerm *whereScanInit( 387 WhereScan *pScan, /* The WhereScan object being initialized */ 388 WhereClause *pWC, /* The WHERE clause to be scanned */ 389 int iCur, /* Cursor to scan for */ 390 int iColumn, /* Column to scan for */ 391 u32 opMask, /* Operator(s) to scan for */ 392 Index *pIdx /* Must be compatible with this index */ 393 ){ 394 pScan->pOrigWC = pWC; 395 pScan->pWC = pWC; 396 pScan->pIdxExpr = 0; 397 pScan->idxaff = 0; 398 pScan->zCollName = 0; 399 pScan->opMask = opMask; 400 pScan->k = 0; 401 pScan->aiCur[0] = iCur; 402 pScan->nEquiv = 1; 403 pScan->iEquiv = 1; 404 if( pIdx ){ 405 int j = iColumn; 406 iColumn = pIdx->aiColumn[j]; 407 if( iColumn==XN_EXPR ){ 408 pScan->pIdxExpr = pIdx->aColExpr->a[j].pExpr; 409 pScan->zCollName = pIdx->azColl[j]; 410 pScan->aiColumn[0] = XN_EXPR; 411 return whereScanInitIndexExpr(pScan); 412 }else if( iColumn==pIdx->pTable->iPKey ){ 413 iColumn = XN_ROWID; 414 }else if( iColumn>=0 ){ 415 pScan->idxaff = pIdx->pTable->aCol[iColumn].affinity; 416 pScan->zCollName = pIdx->azColl[j]; 417 } 418 }else if( iColumn==XN_EXPR ){ 419 return 0; 420 } 421 pScan->aiColumn[0] = iColumn; 422 return whereScanNext(pScan); 423 } 424 425 /* 426 ** Search for a term in the WHERE clause that is of the form "X <op> <expr>" 427 ** where X is a reference to the iColumn of table iCur or of index pIdx 428 ** if pIdx!=0 and <op> is one of the WO_xx operator codes specified by 429 ** the op parameter. Return a pointer to the term. Return 0 if not found. 430 ** 431 ** If pIdx!=0 then it must be one of the indexes of table iCur. 432 ** Search for terms matching the iColumn-th column of pIdx 433 ** rather than the iColumn-th column of table iCur. 434 ** 435 ** The term returned might by Y=<expr> if there is another constraint in 436 ** the WHERE clause that specifies that X=Y. Any such constraints will be 437 ** identified by the WO_EQUIV bit in the pTerm->eOperator field. The 438 ** aiCur[]/iaColumn[] arrays hold X and all its equivalents. There are 11 439 ** slots in aiCur[]/aiColumn[] so that means we can look for X plus up to 10 440 ** other equivalent values. Hence a search for X will return <expr> if X=A1 441 ** and A1=A2 and A2=A3 and ... and A9=A10 and A10=<expr>. 442 ** 443 ** If there are multiple terms in the WHERE clause of the form "X <op> <expr>" 444 ** then try for the one with no dependencies on <expr> - in other words where 445 ** <expr> is a constant expression of some kind. Only return entries of 446 ** the form "X <op> Y" where Y is a column in another table if no terms of 447 ** the form "X <op> <const-expr>" exist. If no terms with a constant RHS 448 ** exist, try to return a term that does not use WO_EQUIV. 449 */ 450 WhereTerm *sqlite3WhereFindTerm( 451 WhereClause *pWC, /* The WHERE clause to be searched */ 452 int iCur, /* Cursor number of LHS */ 453 int iColumn, /* Column number of LHS */ 454 Bitmask notReady, /* RHS must not overlap with this mask */ 455 u32 op, /* Mask of WO_xx values describing operator */ 456 Index *pIdx /* Must be compatible with this index, if not NULL */ 457 ){ 458 WhereTerm *pResult = 0; 459 WhereTerm *p; 460 WhereScan scan; 461 462 p = whereScanInit(&scan, pWC, iCur, iColumn, op, pIdx); 463 op &= WO_EQ|WO_IS; 464 while( p ){ 465 if( (p->prereqRight & notReady)==0 ){ 466 if( p->prereqRight==0 && (p->eOperator&op)!=0 ){ 467 testcase( p->eOperator & WO_IS ); 468 return p; 469 } 470 if( pResult==0 ) pResult = p; 471 } 472 p = whereScanNext(&scan); 473 } 474 return pResult; 475 } 476 477 /* 478 ** This function searches pList for an entry that matches the iCol-th column 479 ** of index pIdx. 480 ** 481 ** If such an expression is found, its index in pList->a[] is returned. If 482 ** no expression is found, -1 is returned. 483 */ 484 static int findIndexCol( 485 Parse *pParse, /* Parse context */ 486 ExprList *pList, /* Expression list to search */ 487 int iBase, /* Cursor for table associated with pIdx */ 488 Index *pIdx, /* Index to match column of */ 489 int iCol /* Column of index to match */ 490 ){ 491 int i; 492 const char *zColl = pIdx->azColl[iCol]; 493 494 for(i=0; i<pList->nExpr; i++){ 495 Expr *p = sqlite3ExprSkipCollateAndLikely(pList->a[i].pExpr); 496 if( ALWAYS(p!=0) 497 && (p->op==TK_COLUMN || p->op==TK_AGG_COLUMN) 498 && p->iColumn==pIdx->aiColumn[iCol] 499 && p->iTable==iBase 500 ){ 501 CollSeq *pColl = sqlite3ExprNNCollSeq(pParse, pList->a[i].pExpr); 502 if( 0==sqlite3StrICmp(pColl->zName, zColl) ){ 503 return i; 504 } 505 } 506 } 507 508 return -1; 509 } 510 511 /* 512 ** Return TRUE if the iCol-th column of index pIdx is NOT NULL 513 */ 514 static int indexColumnNotNull(Index *pIdx, int iCol){ 515 int j; 516 assert( pIdx!=0 ); 517 assert( iCol>=0 && iCol<pIdx->nColumn ); 518 j = pIdx->aiColumn[iCol]; 519 if( j>=0 ){ 520 return pIdx->pTable->aCol[j].notNull; 521 }else if( j==(-1) ){ 522 return 1; 523 }else{ 524 assert( j==(-2) ); 525 return 0; /* Assume an indexed expression can always yield a NULL */ 526 527 } 528 } 529 530 /* 531 ** Return true if the DISTINCT expression-list passed as the third argument 532 ** is redundant. 533 ** 534 ** A DISTINCT list is redundant if any subset of the columns in the 535 ** DISTINCT list are collectively unique and individually non-null. 536 */ 537 static int isDistinctRedundant( 538 Parse *pParse, /* Parsing context */ 539 SrcList *pTabList, /* The FROM clause */ 540 WhereClause *pWC, /* The WHERE clause */ 541 ExprList *pDistinct /* The result set that needs to be DISTINCT */ 542 ){ 543 Table *pTab; 544 Index *pIdx; 545 int i; 546 int iBase; 547 548 /* If there is more than one table or sub-select in the FROM clause of 549 ** this query, then it will not be possible to show that the DISTINCT 550 ** clause is redundant. */ 551 if( pTabList->nSrc!=1 ) return 0; 552 iBase = pTabList->a[0].iCursor; 553 pTab = pTabList->a[0].pTab; 554 555 /* If any of the expressions is an IPK column on table iBase, then return 556 ** true. Note: The (p->iTable==iBase) part of this test may be false if the 557 ** current SELECT is a correlated sub-query. 558 */ 559 for(i=0; i<pDistinct->nExpr; i++){ 560 Expr *p = sqlite3ExprSkipCollateAndLikely(pDistinct->a[i].pExpr); 561 if( NEVER(p==0) ) continue; 562 if( p->op!=TK_COLUMN && p->op!=TK_AGG_COLUMN ) continue; 563 if( p->iTable==iBase && p->iColumn<0 ) return 1; 564 } 565 566 /* Loop through all indices on the table, checking each to see if it makes 567 ** the DISTINCT qualifier redundant. It does so if: 568 ** 569 ** 1. The index is itself UNIQUE, and 570 ** 571 ** 2. All of the columns in the index are either part of the pDistinct 572 ** list, or else the WHERE clause contains a term of the form "col=X", 573 ** where X is a constant value. The collation sequences of the 574 ** comparison and select-list expressions must match those of the index. 575 ** 576 ** 3. All of those index columns for which the WHERE clause does not 577 ** contain a "col=X" term are subject to a NOT NULL constraint. 578 */ 579 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 580 if( !IsUniqueIndex(pIdx) ) continue; 581 if( pIdx->pPartIdxWhere ) continue; 582 for(i=0; i<pIdx->nKeyCol; i++){ 583 if( 0==sqlite3WhereFindTerm(pWC, iBase, i, ~(Bitmask)0, WO_EQ, pIdx) ){ 584 if( findIndexCol(pParse, pDistinct, iBase, pIdx, i)<0 ) break; 585 if( indexColumnNotNull(pIdx, i)==0 ) break; 586 } 587 } 588 if( i==pIdx->nKeyCol ){ 589 /* This index implies that the DISTINCT qualifier is redundant. */ 590 return 1; 591 } 592 } 593 594 return 0; 595 } 596 597 598 /* 599 ** Estimate the logarithm of the input value to base 2. 600 */ 601 static LogEst estLog(LogEst N){ 602 return N<=10 ? 0 : sqlite3LogEst(N) - 33; 603 } 604 605 /* 606 ** Convert OP_Column opcodes to OP_Copy in previously generated code. 607 ** 608 ** This routine runs over generated VDBE code and translates OP_Column 609 ** opcodes into OP_Copy when the table is being accessed via co-routine 610 ** instead of via table lookup. 611 ** 612 ** If the iAutoidxCur is not zero, then any OP_Rowid instructions on 613 ** cursor iTabCur are transformed into OP_Sequence opcode for the 614 ** iAutoidxCur cursor, in order to generate unique rowids for the 615 ** automatic index being generated. 616 */ 617 static void translateColumnToCopy( 618 Parse *pParse, /* Parsing context */ 619 int iStart, /* Translate from this opcode to the end */ 620 int iTabCur, /* OP_Column/OP_Rowid references to this table */ 621 int iRegister, /* The first column is in this register */ 622 int iAutoidxCur /* If non-zero, cursor of autoindex being generated */ 623 ){ 624 Vdbe *v = pParse->pVdbe; 625 VdbeOp *pOp = sqlite3VdbeGetOp(v, iStart); 626 int iEnd = sqlite3VdbeCurrentAddr(v); 627 if( pParse->db->mallocFailed ) return; 628 for(; iStart<iEnd; iStart++, pOp++){ 629 if( pOp->p1!=iTabCur ) continue; 630 if( pOp->opcode==OP_Column ){ 631 pOp->opcode = OP_Copy; 632 pOp->p1 = pOp->p2 + iRegister; 633 pOp->p2 = pOp->p3; 634 pOp->p3 = 0; 635 }else if( pOp->opcode==OP_Rowid ){ 636 pOp->opcode = OP_Sequence; 637 pOp->p1 = iAutoidxCur; 638 #ifdef SQLITE_ALLOW_ROWID_IN_VIEW 639 if( iAutoidxCur==0 ){ 640 pOp->opcode = OP_Null; 641 pOp->p3 = 0; 642 } 643 #endif 644 } 645 } 646 } 647 648 /* 649 ** Two routines for printing the content of an sqlite3_index_info 650 ** structure. Used for testing and debugging only. If neither 651 ** SQLITE_TEST or SQLITE_DEBUG are defined, then these routines 652 ** are no-ops. 653 */ 654 #if !defined(SQLITE_OMIT_VIRTUALTABLE) && defined(WHERETRACE_ENABLED) 655 static void whereTraceIndexInfoInputs(sqlite3_index_info *p){ 656 int i; 657 if( !sqlite3WhereTrace ) return; 658 for(i=0; i<p->nConstraint; i++){ 659 sqlite3DebugPrintf(" constraint[%d]: col=%d termid=%d op=%d usabled=%d\n", 660 i, 661 p->aConstraint[i].iColumn, 662 p->aConstraint[i].iTermOffset, 663 p->aConstraint[i].op, 664 p->aConstraint[i].usable); 665 } 666 for(i=0; i<p->nOrderBy; i++){ 667 sqlite3DebugPrintf(" orderby[%d]: col=%d desc=%d\n", 668 i, 669 p->aOrderBy[i].iColumn, 670 p->aOrderBy[i].desc); 671 } 672 } 673 static void whereTraceIndexInfoOutputs(sqlite3_index_info *p){ 674 int i; 675 if( !sqlite3WhereTrace ) return; 676 for(i=0; i<p->nConstraint; i++){ 677 sqlite3DebugPrintf(" usage[%d]: argvIdx=%d omit=%d\n", 678 i, 679 p->aConstraintUsage[i].argvIndex, 680 p->aConstraintUsage[i].omit); 681 } 682 sqlite3DebugPrintf(" idxNum=%d\n", p->idxNum); 683 sqlite3DebugPrintf(" idxStr=%s\n", p->idxStr); 684 sqlite3DebugPrintf(" orderByConsumed=%d\n", p->orderByConsumed); 685 sqlite3DebugPrintf(" estimatedCost=%g\n", p->estimatedCost); 686 sqlite3DebugPrintf(" estimatedRows=%lld\n", p->estimatedRows); 687 } 688 #else 689 #define whereTraceIndexInfoInputs(A) 690 #define whereTraceIndexInfoOutputs(A) 691 #endif 692 693 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 694 /* 695 ** Return TRUE if the WHERE clause term pTerm is of a form where it 696 ** could be used with an index to access pSrc, assuming an appropriate 697 ** index existed. 698 */ 699 static int termCanDriveIndex( 700 WhereTerm *pTerm, /* WHERE clause term to check */ 701 SrcItem *pSrc, /* Table we are trying to access */ 702 Bitmask notReady /* Tables in outer loops of the join */ 703 ){ 704 char aff; 705 if( pTerm->leftCursor!=pSrc->iCursor ) return 0; 706 if( (pTerm->eOperator & (WO_EQ|WO_IS))==0 ) return 0; 707 if( (pSrc->fg.jointype & JT_LEFT) 708 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 709 && (pTerm->eOperator & WO_IS) 710 ){ 711 /* Cannot use an IS term from the WHERE clause as an index driver for 712 ** the RHS of a LEFT JOIN. Such a term can only be used if it is from 713 ** the ON clause. */ 714 return 0; 715 } 716 if( (pTerm->prereqRight & notReady)!=0 ) return 0; 717 if( pTerm->u.x.leftColumn<0 ) return 0; 718 aff = pSrc->pTab->aCol[pTerm->u.x.leftColumn].affinity; 719 if( !sqlite3IndexAffinityOk(pTerm->pExpr, aff) ) return 0; 720 testcase( pTerm->pExpr->op==TK_IS ); 721 return 1; 722 } 723 #endif 724 725 726 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 727 /* 728 ** Generate code to construct the Index object for an automatic index 729 ** and to set up the WhereLevel object pLevel so that the code generator 730 ** makes use of the automatic index. 731 */ 732 static void constructAutomaticIndex( 733 Parse *pParse, /* The parsing context */ 734 WhereClause *pWC, /* The WHERE clause */ 735 SrcItem *pSrc, /* The FROM clause term to get the next index */ 736 Bitmask notReady, /* Mask of cursors that are not available */ 737 WhereLevel *pLevel /* Write new index here */ 738 ){ 739 int nKeyCol; /* Number of columns in the constructed index */ 740 WhereTerm *pTerm; /* A single term of the WHERE clause */ 741 WhereTerm *pWCEnd; /* End of pWC->a[] */ 742 Index *pIdx; /* Object describing the transient index */ 743 Vdbe *v; /* Prepared statement under construction */ 744 int addrInit; /* Address of the initialization bypass jump */ 745 Table *pTable; /* The table being indexed */ 746 int addrTop; /* Top of the index fill loop */ 747 int regRecord; /* Register holding an index record */ 748 int n; /* Column counter */ 749 int i; /* Loop counter */ 750 int mxBitCol; /* Maximum column in pSrc->colUsed */ 751 CollSeq *pColl; /* Collating sequence to on a column */ 752 WhereLoop *pLoop; /* The Loop object */ 753 char *zNotUsed; /* Extra space on the end of pIdx */ 754 Bitmask idxCols; /* Bitmap of columns used for indexing */ 755 Bitmask extraCols; /* Bitmap of additional columns */ 756 u8 sentWarning = 0; /* True if a warnning has been issued */ 757 Expr *pPartial = 0; /* Partial Index Expression */ 758 int iContinue = 0; /* Jump here to skip excluded rows */ 759 SrcItem *pTabItem; /* FROM clause term being indexed */ 760 int addrCounter = 0; /* Address where integer counter is initialized */ 761 int regBase; /* Array of registers where record is assembled */ 762 763 /* Generate code to skip over the creation and initialization of the 764 ** transient index on 2nd and subsequent iterations of the loop. */ 765 v = pParse->pVdbe; 766 assert( v!=0 ); 767 addrInit = sqlite3VdbeAddOp0(v, OP_Once); VdbeCoverage(v); 768 769 /* Count the number of columns that will be added to the index 770 ** and used to match WHERE clause constraints */ 771 nKeyCol = 0; 772 pTable = pSrc->pTab; 773 pWCEnd = &pWC->a[pWC->nTerm]; 774 pLoop = pLevel->pWLoop; 775 idxCols = 0; 776 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 777 Expr *pExpr = pTerm->pExpr; 778 assert( !ExprHasProperty(pExpr, EP_FromJoin) /* prereq always non-zero */ 779 || pExpr->iRightJoinTable!=pSrc->iCursor /* for the right-hand */ 780 || pLoop->prereq!=0 ); /* table of a LEFT JOIN */ 781 if( pLoop->prereq==0 782 && (pTerm->wtFlags & TERM_VIRTUAL)==0 783 && !ExprHasProperty(pExpr, EP_FromJoin) 784 && sqlite3ExprIsTableConstant(pExpr, pSrc->iCursor) ){ 785 pPartial = sqlite3ExprAnd(pParse, pPartial, 786 sqlite3ExprDup(pParse->db, pExpr, 0)); 787 } 788 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 789 int iCol = pTerm->u.x.leftColumn; 790 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 791 testcase( iCol==BMS ); 792 testcase( iCol==BMS-1 ); 793 if( !sentWarning ){ 794 sqlite3_log(SQLITE_WARNING_AUTOINDEX, 795 "automatic index on %s(%s)", pTable->zName, 796 pTable->aCol[iCol].zName); 797 sentWarning = 1; 798 } 799 if( (idxCols & cMask)==0 ){ 800 if( whereLoopResize(pParse->db, pLoop, nKeyCol+1) ){ 801 goto end_auto_index_create; 802 } 803 pLoop->aLTerm[nKeyCol++] = pTerm; 804 idxCols |= cMask; 805 } 806 } 807 } 808 assert( nKeyCol>0 || pParse->db->mallocFailed ); 809 pLoop->u.btree.nEq = pLoop->nLTerm = nKeyCol; 810 pLoop->wsFlags = WHERE_COLUMN_EQ | WHERE_IDX_ONLY | WHERE_INDEXED 811 | WHERE_AUTO_INDEX; 812 813 /* Count the number of additional columns needed to create a 814 ** covering index. A "covering index" is an index that contains all 815 ** columns that are needed by the query. With a covering index, the 816 ** original table never needs to be accessed. Automatic indices must 817 ** be a covering index because the index will not be updated if the 818 ** original table changes and the index and table cannot both be used 819 ** if they go out of sync. 820 */ 821 extraCols = pSrc->colUsed & (~idxCols | MASKBIT(BMS-1)); 822 mxBitCol = MIN(BMS-1,pTable->nCol); 823 testcase( pTable->nCol==BMS-1 ); 824 testcase( pTable->nCol==BMS-2 ); 825 for(i=0; i<mxBitCol; i++){ 826 if( extraCols & MASKBIT(i) ) nKeyCol++; 827 } 828 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 829 nKeyCol += pTable->nCol - BMS + 1; 830 } 831 832 /* Construct the Index object to describe this index */ 833 pIdx = sqlite3AllocateIndexObject(pParse->db, nKeyCol+1, 0, &zNotUsed); 834 if( pIdx==0 ) goto end_auto_index_create; 835 pLoop->u.btree.pIndex = pIdx; 836 pIdx->zName = "auto-index"; 837 pIdx->pTable = pTable; 838 n = 0; 839 idxCols = 0; 840 for(pTerm=pWC->a; pTerm<pWCEnd; pTerm++){ 841 if( termCanDriveIndex(pTerm, pSrc, notReady) ){ 842 int iCol = pTerm->u.x.leftColumn; 843 Bitmask cMask = iCol>=BMS ? MASKBIT(BMS-1) : MASKBIT(iCol); 844 testcase( iCol==BMS-1 ); 845 testcase( iCol==BMS ); 846 if( (idxCols & cMask)==0 ){ 847 Expr *pX = pTerm->pExpr; 848 idxCols |= cMask; 849 pIdx->aiColumn[n] = pTerm->u.x.leftColumn; 850 pColl = sqlite3ExprCompareCollSeq(pParse, pX); 851 assert( pColl!=0 || pParse->nErr>0 ); /* TH3 collate01.800 */ 852 pIdx->azColl[n] = pColl ? pColl->zName : sqlite3StrBINARY; 853 n++; 854 } 855 } 856 } 857 assert( (u32)n==pLoop->u.btree.nEq ); 858 859 /* Add additional columns needed to make the automatic index into 860 ** a covering index */ 861 for(i=0; i<mxBitCol; i++){ 862 if( extraCols & MASKBIT(i) ){ 863 pIdx->aiColumn[n] = i; 864 pIdx->azColl[n] = sqlite3StrBINARY; 865 n++; 866 } 867 } 868 if( pSrc->colUsed & MASKBIT(BMS-1) ){ 869 for(i=BMS-1; i<pTable->nCol; i++){ 870 pIdx->aiColumn[n] = i; 871 pIdx->azColl[n] = sqlite3StrBINARY; 872 n++; 873 } 874 } 875 assert( n==nKeyCol ); 876 pIdx->aiColumn[n] = XN_ROWID; 877 pIdx->azColl[n] = sqlite3StrBINARY; 878 879 /* Create the automatic index */ 880 assert( pLevel->iIdxCur>=0 ); 881 pLevel->iIdxCur = pParse->nTab++; 882 sqlite3VdbeAddOp2(v, OP_OpenAutoindex, pLevel->iIdxCur, nKeyCol+1); 883 sqlite3VdbeSetP4KeyInfo(pParse, pIdx); 884 VdbeComment((v, "for %s", pTable->zName)); 885 886 /* Fill the automatic index with content */ 887 pTabItem = &pWC->pWInfo->pTabList->a[pLevel->iFrom]; 888 if( pTabItem->fg.viaCoroutine ){ 889 int regYield = pTabItem->regReturn; 890 addrCounter = sqlite3VdbeAddOp2(v, OP_Integer, 0, 0); 891 sqlite3VdbeAddOp3(v, OP_InitCoroutine, regYield, 0, pTabItem->addrFillSub); 892 addrTop = sqlite3VdbeAddOp1(v, OP_Yield, regYield); 893 VdbeCoverage(v); 894 VdbeComment((v, "next row of %s", pTabItem->pTab->zName)); 895 }else{ 896 addrTop = sqlite3VdbeAddOp1(v, OP_Rewind, pLevel->iTabCur); VdbeCoverage(v); 897 } 898 if( pPartial ){ 899 iContinue = sqlite3VdbeMakeLabel(pParse); 900 sqlite3ExprIfFalse(pParse, pPartial, iContinue, SQLITE_JUMPIFNULL); 901 pLoop->wsFlags |= WHERE_PARTIALIDX; 902 } 903 regRecord = sqlite3GetTempReg(pParse); 904 regBase = sqlite3GenerateIndexKey( 905 pParse, pIdx, pLevel->iTabCur, regRecord, 0, 0, 0, 0 906 ); 907 sqlite3VdbeAddOp2(v, OP_IdxInsert, pLevel->iIdxCur, regRecord); 908 sqlite3VdbeChangeP5(v, OPFLAG_USESEEKRESULT); 909 if( pPartial ) sqlite3VdbeResolveLabel(v, iContinue); 910 if( pTabItem->fg.viaCoroutine ){ 911 sqlite3VdbeChangeP2(v, addrCounter, regBase+n); 912 testcase( pParse->db->mallocFailed ); 913 assert( pLevel->iIdxCur>0 ); 914 translateColumnToCopy(pParse, addrTop, pLevel->iTabCur, 915 pTabItem->regResult, pLevel->iIdxCur); 916 sqlite3VdbeGoto(v, addrTop); 917 pTabItem->fg.viaCoroutine = 0; 918 }else{ 919 sqlite3VdbeAddOp2(v, OP_Next, pLevel->iTabCur, addrTop+1); VdbeCoverage(v); 920 sqlite3VdbeChangeP5(v, SQLITE_STMTSTATUS_AUTOINDEX); 921 } 922 sqlite3VdbeJumpHere(v, addrTop); 923 sqlite3ReleaseTempReg(pParse, regRecord); 924 925 /* Jump here when skipping the initialization */ 926 sqlite3VdbeJumpHere(v, addrInit); 927 928 end_auto_index_create: 929 sqlite3ExprDelete(pParse->db, pPartial); 930 } 931 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 932 933 #ifndef SQLITE_OMIT_VIRTUALTABLE 934 /* 935 ** Allocate and populate an sqlite3_index_info structure. It is the 936 ** responsibility of the caller to eventually release the structure 937 ** by passing the pointer returned by this function to sqlite3_free(). 938 */ 939 static sqlite3_index_info *allocateIndexInfo( 940 Parse *pParse, /* The parsing context */ 941 WhereClause *pWC, /* The WHERE clause being analyzed */ 942 Bitmask mUnusable, /* Ignore terms with these prereqs */ 943 SrcItem *pSrc, /* The FROM clause term that is the vtab */ 944 ExprList *pOrderBy, /* The ORDER BY clause */ 945 u16 *pmNoOmit /* Mask of terms not to omit */ 946 ){ 947 int i, j; 948 int nTerm; 949 struct sqlite3_index_constraint *pIdxCons; 950 struct sqlite3_index_orderby *pIdxOrderBy; 951 struct sqlite3_index_constraint_usage *pUsage; 952 struct HiddenIndexInfo *pHidden; 953 WhereTerm *pTerm; 954 int nOrderBy; 955 sqlite3_index_info *pIdxInfo; 956 u16 mNoOmit = 0; 957 958 /* Count the number of possible WHERE clause constraints referring 959 ** to this virtual table */ 960 for(i=nTerm=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 961 if( pTerm->leftCursor != pSrc->iCursor ) continue; 962 if( pTerm->prereqRight & mUnusable ) continue; 963 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 964 testcase( pTerm->eOperator & WO_IN ); 965 testcase( pTerm->eOperator & WO_ISNULL ); 966 testcase( pTerm->eOperator & WO_IS ); 967 testcase( pTerm->eOperator & WO_ALL ); 968 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 969 if( pTerm->wtFlags & TERM_VNULL ) continue; 970 assert( pTerm->u.x.leftColumn>=(-1) ); 971 nTerm++; 972 } 973 974 /* If the ORDER BY clause contains only columns in the current 975 ** virtual table then allocate space for the aOrderBy part of 976 ** the sqlite3_index_info structure. 977 */ 978 nOrderBy = 0; 979 if( pOrderBy ){ 980 int n = pOrderBy->nExpr; 981 for(i=0; i<n; i++){ 982 Expr *pExpr = pOrderBy->a[i].pExpr; 983 if( pExpr->op!=TK_COLUMN || pExpr->iTable!=pSrc->iCursor ) break; 984 if( pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL ) break; 985 } 986 if( i==n){ 987 nOrderBy = n; 988 } 989 } 990 991 /* Allocate the sqlite3_index_info structure 992 */ 993 pIdxInfo = sqlite3DbMallocZero(pParse->db, sizeof(*pIdxInfo) 994 + (sizeof(*pIdxCons) + sizeof(*pUsage))*nTerm 995 + sizeof(*pIdxOrderBy)*nOrderBy + sizeof(*pHidden) ); 996 if( pIdxInfo==0 ){ 997 sqlite3ErrorMsg(pParse, "out of memory"); 998 return 0; 999 } 1000 pHidden = (struct HiddenIndexInfo*)&pIdxInfo[1]; 1001 pIdxCons = (struct sqlite3_index_constraint*)&pHidden[1]; 1002 pIdxOrderBy = (struct sqlite3_index_orderby*)&pIdxCons[nTerm]; 1003 pUsage = (struct sqlite3_index_constraint_usage*)&pIdxOrderBy[nOrderBy]; 1004 pIdxInfo->nOrderBy = nOrderBy; 1005 pIdxInfo->aConstraint = pIdxCons; 1006 pIdxInfo->aOrderBy = pIdxOrderBy; 1007 pIdxInfo->aConstraintUsage = pUsage; 1008 pHidden->pWC = pWC; 1009 pHidden->pParse = pParse; 1010 for(i=j=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 1011 u16 op; 1012 if( pTerm->leftCursor != pSrc->iCursor ) continue; 1013 if( pTerm->prereqRight & mUnusable ) continue; 1014 assert( IsPowerOfTwo(pTerm->eOperator & ~WO_EQUIV) ); 1015 testcase( pTerm->eOperator & WO_IN ); 1016 testcase( pTerm->eOperator & WO_IS ); 1017 testcase( pTerm->eOperator & WO_ISNULL ); 1018 testcase( pTerm->eOperator & WO_ALL ); 1019 if( (pTerm->eOperator & ~(WO_EQUIV))==0 ) continue; 1020 if( pTerm->wtFlags & TERM_VNULL ) continue; 1021 1022 /* tag-20191211-002: WHERE-clause constraints are not useful to the 1023 ** right-hand table of a LEFT JOIN. See tag-20191211-001 for the 1024 ** equivalent restriction for ordinary tables. */ 1025 if( (pSrc->fg.jointype & JT_LEFT)!=0 1026 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 1027 ){ 1028 continue; 1029 } 1030 assert( pTerm->u.x.leftColumn>=(-1) ); 1031 pIdxCons[j].iColumn = pTerm->u.x.leftColumn; 1032 pIdxCons[j].iTermOffset = i; 1033 op = pTerm->eOperator & WO_ALL; 1034 if( op==WO_IN ) op = WO_EQ; 1035 if( op==WO_AUX ){ 1036 pIdxCons[j].op = pTerm->eMatchOp; 1037 }else if( op & (WO_ISNULL|WO_IS) ){ 1038 if( op==WO_ISNULL ){ 1039 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_ISNULL; 1040 }else{ 1041 pIdxCons[j].op = SQLITE_INDEX_CONSTRAINT_IS; 1042 } 1043 }else{ 1044 pIdxCons[j].op = (u8)op; 1045 /* The direct assignment in the previous line is possible only because 1046 ** the WO_ and SQLITE_INDEX_CONSTRAINT_ codes are identical. The 1047 ** following asserts verify this fact. */ 1048 assert( WO_EQ==SQLITE_INDEX_CONSTRAINT_EQ ); 1049 assert( WO_LT==SQLITE_INDEX_CONSTRAINT_LT ); 1050 assert( WO_LE==SQLITE_INDEX_CONSTRAINT_LE ); 1051 assert( WO_GT==SQLITE_INDEX_CONSTRAINT_GT ); 1052 assert( WO_GE==SQLITE_INDEX_CONSTRAINT_GE ); 1053 assert( pTerm->eOperator&(WO_IN|WO_EQ|WO_LT|WO_LE|WO_GT|WO_GE|WO_AUX) ); 1054 1055 if( op & (WO_LT|WO_LE|WO_GT|WO_GE) 1056 && sqlite3ExprIsVector(pTerm->pExpr->pRight) 1057 ){ 1058 testcase( j!=i ); 1059 if( j<16 ) mNoOmit |= (1 << j); 1060 if( op==WO_LT ) pIdxCons[j].op = WO_LE; 1061 if( op==WO_GT ) pIdxCons[j].op = WO_GE; 1062 } 1063 } 1064 1065 j++; 1066 } 1067 pIdxInfo->nConstraint = j; 1068 for(i=0; i<nOrderBy; i++){ 1069 Expr *pExpr = pOrderBy->a[i].pExpr; 1070 pIdxOrderBy[i].iColumn = pExpr->iColumn; 1071 pIdxOrderBy[i].desc = pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC; 1072 } 1073 1074 *pmNoOmit = mNoOmit; 1075 return pIdxInfo; 1076 } 1077 1078 /* 1079 ** The table object reference passed as the second argument to this function 1080 ** must represent a virtual table. This function invokes the xBestIndex() 1081 ** method of the virtual table with the sqlite3_index_info object that 1082 ** comes in as the 3rd argument to this function. 1083 ** 1084 ** If an error occurs, pParse is populated with an error message and an 1085 ** appropriate error code is returned. A return of SQLITE_CONSTRAINT from 1086 ** xBestIndex is not considered an error. SQLITE_CONSTRAINT indicates that 1087 ** the current configuration of "unusable" flags in sqlite3_index_info can 1088 ** not result in a valid plan. 1089 ** 1090 ** Whether or not an error is returned, it is the responsibility of the 1091 ** caller to eventually free p->idxStr if p->needToFreeIdxStr indicates 1092 ** that this is required. 1093 */ 1094 static int vtabBestIndex(Parse *pParse, Table *pTab, sqlite3_index_info *p){ 1095 sqlite3_vtab *pVtab = sqlite3GetVTable(pParse->db, pTab)->pVtab; 1096 int rc; 1097 1098 whereTraceIndexInfoInputs(p); 1099 rc = pVtab->pModule->xBestIndex(pVtab, p); 1100 whereTraceIndexInfoOutputs(p); 1101 1102 if( rc!=SQLITE_OK && rc!=SQLITE_CONSTRAINT ){ 1103 if( rc==SQLITE_NOMEM ){ 1104 sqlite3OomFault(pParse->db); 1105 }else if( !pVtab->zErrMsg ){ 1106 sqlite3ErrorMsg(pParse, "%s", sqlite3ErrStr(rc)); 1107 }else{ 1108 sqlite3ErrorMsg(pParse, "%s", pVtab->zErrMsg); 1109 } 1110 } 1111 sqlite3_free(pVtab->zErrMsg); 1112 pVtab->zErrMsg = 0; 1113 return rc; 1114 } 1115 #endif /* !defined(SQLITE_OMIT_VIRTUALTABLE) */ 1116 1117 #ifdef SQLITE_ENABLE_STAT4 1118 /* 1119 ** Estimate the location of a particular key among all keys in an 1120 ** index. Store the results in aStat as follows: 1121 ** 1122 ** aStat[0] Est. number of rows less than pRec 1123 ** aStat[1] Est. number of rows equal to pRec 1124 ** 1125 ** Return the index of the sample that is the smallest sample that 1126 ** is greater than or equal to pRec. Note that this index is not an index 1127 ** into the aSample[] array - it is an index into a virtual set of samples 1128 ** based on the contents of aSample[] and the number of fields in record 1129 ** pRec. 1130 */ 1131 static int whereKeyStats( 1132 Parse *pParse, /* Database connection */ 1133 Index *pIdx, /* Index to consider domain of */ 1134 UnpackedRecord *pRec, /* Vector of values to consider */ 1135 int roundUp, /* Round up if true. Round down if false */ 1136 tRowcnt *aStat /* OUT: stats written here */ 1137 ){ 1138 IndexSample *aSample = pIdx->aSample; 1139 int iCol; /* Index of required stats in anEq[] etc. */ 1140 int i; /* Index of first sample >= pRec */ 1141 int iSample; /* Smallest sample larger than or equal to pRec */ 1142 int iMin = 0; /* Smallest sample not yet tested */ 1143 int iTest; /* Next sample to test */ 1144 int res; /* Result of comparison operation */ 1145 int nField; /* Number of fields in pRec */ 1146 tRowcnt iLower = 0; /* anLt[] + anEq[] of largest sample pRec is > */ 1147 1148 #ifndef SQLITE_DEBUG 1149 UNUSED_PARAMETER( pParse ); 1150 #endif 1151 assert( pRec!=0 ); 1152 assert( pIdx->nSample>0 ); 1153 assert( pRec->nField>0 && pRec->nField<=pIdx->nSampleCol ); 1154 1155 /* Do a binary search to find the first sample greater than or equal 1156 ** to pRec. If pRec contains a single field, the set of samples to search 1157 ** is simply the aSample[] array. If the samples in aSample[] contain more 1158 ** than one fields, all fields following the first are ignored. 1159 ** 1160 ** If pRec contains N fields, where N is more than one, then as well as the 1161 ** samples in aSample[] (truncated to N fields), the search also has to 1162 ** consider prefixes of those samples. For example, if the set of samples 1163 ** in aSample is: 1164 ** 1165 ** aSample[0] = (a, 5) 1166 ** aSample[1] = (a, 10) 1167 ** aSample[2] = (b, 5) 1168 ** aSample[3] = (c, 100) 1169 ** aSample[4] = (c, 105) 1170 ** 1171 ** Then the search space should ideally be the samples above and the 1172 ** unique prefixes [a], [b] and [c]. But since that is hard to organize, 1173 ** the code actually searches this set: 1174 ** 1175 ** 0: (a) 1176 ** 1: (a, 5) 1177 ** 2: (a, 10) 1178 ** 3: (a, 10) 1179 ** 4: (b) 1180 ** 5: (b, 5) 1181 ** 6: (c) 1182 ** 7: (c, 100) 1183 ** 8: (c, 105) 1184 ** 9: (c, 105) 1185 ** 1186 ** For each sample in the aSample[] array, N samples are present in the 1187 ** effective sample array. In the above, samples 0 and 1 are based on 1188 ** sample aSample[0]. Samples 2 and 3 on aSample[1] etc. 1189 ** 1190 ** Often, sample i of each block of N effective samples has (i+1) fields. 1191 ** Except, each sample may be extended to ensure that it is greater than or 1192 ** equal to the previous sample in the array. For example, in the above, 1193 ** sample 2 is the first sample of a block of N samples, so at first it 1194 ** appears that it should be 1 field in size. However, that would make it 1195 ** smaller than sample 1, so the binary search would not work. As a result, 1196 ** it is extended to two fields. The duplicates that this creates do not 1197 ** cause any problems. 1198 */ 1199 nField = pRec->nField; 1200 iCol = 0; 1201 iSample = pIdx->nSample * nField; 1202 do{ 1203 int iSamp; /* Index in aSample[] of test sample */ 1204 int n; /* Number of fields in test sample */ 1205 1206 iTest = (iMin+iSample)/2; 1207 iSamp = iTest / nField; 1208 if( iSamp>0 ){ 1209 /* The proposed effective sample is a prefix of sample aSample[iSamp]. 1210 ** Specifically, the shortest prefix of at least (1 + iTest%nField) 1211 ** fields that is greater than the previous effective sample. */ 1212 for(n=(iTest % nField) + 1; n<nField; n++){ 1213 if( aSample[iSamp-1].anLt[n-1]!=aSample[iSamp].anLt[n-1] ) break; 1214 } 1215 }else{ 1216 n = iTest + 1; 1217 } 1218 1219 pRec->nField = n; 1220 res = sqlite3VdbeRecordCompare(aSample[iSamp].n, aSample[iSamp].p, pRec); 1221 if( res<0 ){ 1222 iLower = aSample[iSamp].anLt[n-1] + aSample[iSamp].anEq[n-1]; 1223 iMin = iTest+1; 1224 }else if( res==0 && n<nField ){ 1225 iLower = aSample[iSamp].anLt[n-1]; 1226 iMin = iTest+1; 1227 res = -1; 1228 }else{ 1229 iSample = iTest; 1230 iCol = n-1; 1231 } 1232 }while( res && iMin<iSample ); 1233 i = iSample / nField; 1234 1235 #ifdef SQLITE_DEBUG 1236 /* The following assert statements check that the binary search code 1237 ** above found the right answer. This block serves no purpose other 1238 ** than to invoke the asserts. */ 1239 if( pParse->db->mallocFailed==0 ){ 1240 if( res==0 ){ 1241 /* If (res==0) is true, then pRec must be equal to sample i. */ 1242 assert( i<pIdx->nSample ); 1243 assert( iCol==nField-1 ); 1244 pRec->nField = nField; 1245 assert( 0==sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec) 1246 || pParse->db->mallocFailed 1247 ); 1248 }else{ 1249 /* Unless i==pIdx->nSample, indicating that pRec is larger than 1250 ** all samples in the aSample[] array, pRec must be smaller than the 1251 ** (iCol+1) field prefix of sample i. */ 1252 assert( i<=pIdx->nSample && i>=0 ); 1253 pRec->nField = iCol+1; 1254 assert( i==pIdx->nSample 1255 || sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)>0 1256 || pParse->db->mallocFailed ); 1257 1258 /* if i==0 and iCol==0, then record pRec is smaller than all samples 1259 ** in the aSample[] array. Otherwise, if (iCol>0) then pRec must 1260 ** be greater than or equal to the (iCol) field prefix of sample i. 1261 ** If (i>0), then pRec must also be greater than sample (i-1). */ 1262 if( iCol>0 ){ 1263 pRec->nField = iCol; 1264 assert( sqlite3VdbeRecordCompare(aSample[i].n, aSample[i].p, pRec)<=0 1265 || pParse->db->mallocFailed ); 1266 } 1267 if( i>0 ){ 1268 pRec->nField = nField; 1269 assert( sqlite3VdbeRecordCompare(aSample[i-1].n, aSample[i-1].p, pRec)<0 1270 || pParse->db->mallocFailed ); 1271 } 1272 } 1273 } 1274 #endif /* ifdef SQLITE_DEBUG */ 1275 1276 if( res==0 ){ 1277 /* Record pRec is equal to sample i */ 1278 assert( iCol==nField-1 ); 1279 aStat[0] = aSample[i].anLt[iCol]; 1280 aStat[1] = aSample[i].anEq[iCol]; 1281 }else{ 1282 /* At this point, the (iCol+1) field prefix of aSample[i] is the first 1283 ** sample that is greater than pRec. Or, if i==pIdx->nSample then pRec 1284 ** is larger than all samples in the array. */ 1285 tRowcnt iUpper, iGap; 1286 if( i>=pIdx->nSample ){ 1287 iUpper = sqlite3LogEstToInt(pIdx->aiRowLogEst[0]); 1288 }else{ 1289 iUpper = aSample[i].anLt[iCol]; 1290 } 1291 1292 if( iLower>=iUpper ){ 1293 iGap = 0; 1294 }else{ 1295 iGap = iUpper - iLower; 1296 } 1297 if( roundUp ){ 1298 iGap = (iGap*2)/3; 1299 }else{ 1300 iGap = iGap/3; 1301 } 1302 aStat[0] = iLower + iGap; 1303 aStat[1] = pIdx->aAvgEq[nField-1]; 1304 } 1305 1306 /* Restore the pRec->nField value before returning. */ 1307 pRec->nField = nField; 1308 return i; 1309 } 1310 #endif /* SQLITE_ENABLE_STAT4 */ 1311 1312 /* 1313 ** If it is not NULL, pTerm is a term that provides an upper or lower 1314 ** bound on a range scan. Without considering pTerm, it is estimated 1315 ** that the scan will visit nNew rows. This function returns the number 1316 ** estimated to be visited after taking pTerm into account. 1317 ** 1318 ** If the user explicitly specified a likelihood() value for this term, 1319 ** then the return value is the likelihood multiplied by the number of 1320 ** input rows. Otherwise, this function assumes that an "IS NOT NULL" term 1321 ** has a likelihood of 0.50, and any other term a likelihood of 0.25. 1322 */ 1323 static LogEst whereRangeAdjust(WhereTerm *pTerm, LogEst nNew){ 1324 LogEst nRet = nNew; 1325 if( pTerm ){ 1326 if( pTerm->truthProb<=0 ){ 1327 nRet += pTerm->truthProb; 1328 }else if( (pTerm->wtFlags & TERM_VNULL)==0 ){ 1329 nRet -= 20; assert( 20==sqlite3LogEst(4) ); 1330 } 1331 } 1332 return nRet; 1333 } 1334 1335 1336 #ifdef SQLITE_ENABLE_STAT4 1337 /* 1338 ** Return the affinity for a single column of an index. 1339 */ 1340 char sqlite3IndexColumnAffinity(sqlite3 *db, Index *pIdx, int iCol){ 1341 assert( iCol>=0 && iCol<pIdx->nColumn ); 1342 if( !pIdx->zColAff ){ 1343 if( sqlite3IndexAffinityStr(db, pIdx)==0 ) return SQLITE_AFF_BLOB; 1344 } 1345 assert( pIdx->zColAff[iCol]!=0 ); 1346 return pIdx->zColAff[iCol]; 1347 } 1348 #endif 1349 1350 1351 #ifdef SQLITE_ENABLE_STAT4 1352 /* 1353 ** This function is called to estimate the number of rows visited by a 1354 ** range-scan on a skip-scan index. For example: 1355 ** 1356 ** CREATE INDEX i1 ON t1(a, b, c); 1357 ** SELECT * FROM t1 WHERE a=? AND c BETWEEN ? AND ?; 1358 ** 1359 ** Value pLoop->nOut is currently set to the estimated number of rows 1360 ** visited for scanning (a=? AND b=?). This function reduces that estimate 1361 ** by some factor to account for the (c BETWEEN ? AND ?) expression based 1362 ** on the stat4 data for the index. this scan will be peformed multiple 1363 ** times (once for each (a,b) combination that matches a=?) is dealt with 1364 ** by the caller. 1365 ** 1366 ** It does this by scanning through all stat4 samples, comparing values 1367 ** extracted from pLower and pUpper with the corresponding column in each 1368 ** sample. If L and U are the number of samples found to be less than or 1369 ** equal to the values extracted from pLower and pUpper respectively, and 1370 ** N is the total number of samples, the pLoop->nOut value is adjusted 1371 ** as follows: 1372 ** 1373 ** nOut = nOut * ( min(U - L, 1) / N ) 1374 ** 1375 ** If pLower is NULL, or a value cannot be extracted from the term, L is 1376 ** set to zero. If pUpper is NULL, or a value cannot be extracted from it, 1377 ** U is set to N. 1378 ** 1379 ** Normally, this function sets *pbDone to 1 before returning. However, 1380 ** if no value can be extracted from either pLower or pUpper (and so the 1381 ** estimate of the number of rows delivered remains unchanged), *pbDone 1382 ** is left as is. 1383 ** 1384 ** If an error occurs, an SQLite error code is returned. Otherwise, 1385 ** SQLITE_OK. 1386 */ 1387 static int whereRangeSkipScanEst( 1388 Parse *pParse, /* Parsing & code generating context */ 1389 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1390 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1391 WhereLoop *pLoop, /* Update the .nOut value of this loop */ 1392 int *pbDone /* Set to true if at least one expr. value extracted */ 1393 ){ 1394 Index *p = pLoop->u.btree.pIndex; 1395 int nEq = pLoop->u.btree.nEq; 1396 sqlite3 *db = pParse->db; 1397 int nLower = -1; 1398 int nUpper = p->nSample+1; 1399 int rc = SQLITE_OK; 1400 u8 aff = sqlite3IndexColumnAffinity(db, p, nEq); 1401 CollSeq *pColl; 1402 1403 sqlite3_value *p1 = 0; /* Value extracted from pLower */ 1404 sqlite3_value *p2 = 0; /* Value extracted from pUpper */ 1405 sqlite3_value *pVal = 0; /* Value extracted from record */ 1406 1407 pColl = sqlite3LocateCollSeq(pParse, p->azColl[nEq]); 1408 if( pLower ){ 1409 rc = sqlite3Stat4ValueFromExpr(pParse, pLower->pExpr->pRight, aff, &p1); 1410 nLower = 0; 1411 } 1412 if( pUpper && rc==SQLITE_OK ){ 1413 rc = sqlite3Stat4ValueFromExpr(pParse, pUpper->pExpr->pRight, aff, &p2); 1414 nUpper = p2 ? 0 : p->nSample; 1415 } 1416 1417 if( p1 || p2 ){ 1418 int i; 1419 int nDiff; 1420 for(i=0; rc==SQLITE_OK && i<p->nSample; i++){ 1421 rc = sqlite3Stat4Column(db, p->aSample[i].p, p->aSample[i].n, nEq, &pVal); 1422 if( rc==SQLITE_OK && p1 ){ 1423 int res = sqlite3MemCompare(p1, pVal, pColl); 1424 if( res>=0 ) nLower++; 1425 } 1426 if( rc==SQLITE_OK && p2 ){ 1427 int res = sqlite3MemCompare(p2, pVal, pColl); 1428 if( res>=0 ) nUpper++; 1429 } 1430 } 1431 nDiff = (nUpper - nLower); 1432 if( nDiff<=0 ) nDiff = 1; 1433 1434 /* If there is both an upper and lower bound specified, and the 1435 ** comparisons indicate that they are close together, use the fallback 1436 ** method (assume that the scan visits 1/64 of the rows) for estimating 1437 ** the number of rows visited. Otherwise, estimate the number of rows 1438 ** using the method described in the header comment for this function. */ 1439 if( nDiff!=1 || pUpper==0 || pLower==0 ){ 1440 int nAdjust = (sqlite3LogEst(p->nSample) - sqlite3LogEst(nDiff)); 1441 pLoop->nOut -= nAdjust; 1442 *pbDone = 1; 1443 WHERETRACE(0x10, ("range skip-scan regions: %u..%u adjust=%d est=%d\n", 1444 nLower, nUpper, nAdjust*-1, pLoop->nOut)); 1445 } 1446 1447 }else{ 1448 assert( *pbDone==0 ); 1449 } 1450 1451 sqlite3ValueFree(p1); 1452 sqlite3ValueFree(p2); 1453 sqlite3ValueFree(pVal); 1454 1455 return rc; 1456 } 1457 #endif /* SQLITE_ENABLE_STAT4 */ 1458 1459 /* 1460 ** This function is used to estimate the number of rows that will be visited 1461 ** by scanning an index for a range of values. The range may have an upper 1462 ** bound, a lower bound, or both. The WHERE clause terms that set the upper 1463 ** and lower bounds are represented by pLower and pUpper respectively. For 1464 ** example, assuming that index p is on t1(a): 1465 ** 1466 ** ... FROM t1 WHERE a > ? AND a < ? ... 1467 ** |_____| |_____| 1468 ** | | 1469 ** pLower pUpper 1470 ** 1471 ** If either of the upper or lower bound is not present, then NULL is passed in 1472 ** place of the corresponding WhereTerm. 1473 ** 1474 ** The value in (pBuilder->pNew->u.btree.nEq) is the number of the index 1475 ** column subject to the range constraint. Or, equivalently, the number of 1476 ** equality constraints optimized by the proposed index scan. For example, 1477 ** assuming index p is on t1(a, b), and the SQL query is: 1478 ** 1479 ** ... FROM t1 WHERE a = ? AND b > ? AND b < ? ... 1480 ** 1481 ** then nEq is set to 1 (as the range restricted column, b, is the second 1482 ** left-most column of the index). Or, if the query is: 1483 ** 1484 ** ... FROM t1 WHERE a > ? AND a < ? ... 1485 ** 1486 ** then nEq is set to 0. 1487 ** 1488 ** When this function is called, *pnOut is set to the sqlite3LogEst() of the 1489 ** number of rows that the index scan is expected to visit without 1490 ** considering the range constraints. If nEq is 0, then *pnOut is the number of 1491 ** rows in the index. Assuming no error occurs, *pnOut is adjusted (reduced) 1492 ** to account for the range constraints pLower and pUpper. 1493 ** 1494 ** In the absence of sqlite_stat4 ANALYZE data, or if such data cannot be 1495 ** used, a single range inequality reduces the search space by a factor of 4. 1496 ** and a pair of constraints (x>? AND x<?) reduces the expected number of 1497 ** rows visited by a factor of 64. 1498 */ 1499 static int whereRangeScanEst( 1500 Parse *pParse, /* Parsing & code generating context */ 1501 WhereLoopBuilder *pBuilder, 1502 WhereTerm *pLower, /* Lower bound on the range. ex: "x>123" Might be NULL */ 1503 WhereTerm *pUpper, /* Upper bound on the range. ex: "x<455" Might be NULL */ 1504 WhereLoop *pLoop /* Modify the .nOut and maybe .rRun fields */ 1505 ){ 1506 int rc = SQLITE_OK; 1507 int nOut = pLoop->nOut; 1508 LogEst nNew; 1509 1510 #ifdef SQLITE_ENABLE_STAT4 1511 Index *p = pLoop->u.btree.pIndex; 1512 int nEq = pLoop->u.btree.nEq; 1513 1514 if( p->nSample>0 && ALWAYS(nEq<p->nSampleCol) 1515 && OptimizationEnabled(pParse->db, SQLITE_Stat4) 1516 ){ 1517 if( nEq==pBuilder->nRecValid ){ 1518 UnpackedRecord *pRec = pBuilder->pRec; 1519 tRowcnt a[2]; 1520 int nBtm = pLoop->u.btree.nBtm; 1521 int nTop = pLoop->u.btree.nTop; 1522 1523 /* Variable iLower will be set to the estimate of the number of rows in 1524 ** the index that are less than the lower bound of the range query. The 1525 ** lower bound being the concatenation of $P and $L, where $P is the 1526 ** key-prefix formed by the nEq values matched against the nEq left-most 1527 ** columns of the index, and $L is the value in pLower. 1528 ** 1529 ** Or, if pLower is NULL or $L cannot be extracted from it (because it 1530 ** is not a simple variable or literal value), the lower bound of the 1531 ** range is $P. Due to a quirk in the way whereKeyStats() works, even 1532 ** if $L is available, whereKeyStats() is called for both ($P) and 1533 ** ($P:$L) and the larger of the two returned values is used. 1534 ** 1535 ** Similarly, iUpper is to be set to the estimate of the number of rows 1536 ** less than the upper bound of the range query. Where the upper bound 1537 ** is either ($P) or ($P:$U). Again, even if $U is available, both values 1538 ** of iUpper are requested of whereKeyStats() and the smaller used. 1539 ** 1540 ** The number of rows between the two bounds is then just iUpper-iLower. 1541 */ 1542 tRowcnt iLower; /* Rows less than the lower bound */ 1543 tRowcnt iUpper; /* Rows less than the upper bound */ 1544 int iLwrIdx = -2; /* aSample[] for the lower bound */ 1545 int iUprIdx = -1; /* aSample[] for the upper bound */ 1546 1547 if( pRec ){ 1548 testcase( pRec->nField!=pBuilder->nRecValid ); 1549 pRec->nField = pBuilder->nRecValid; 1550 } 1551 /* Determine iLower and iUpper using ($P) only. */ 1552 if( nEq==0 ){ 1553 iLower = 0; 1554 iUpper = p->nRowEst0; 1555 }else{ 1556 /* Note: this call could be optimized away - since the same values must 1557 ** have been requested when testing key $P in whereEqualScanEst(). */ 1558 whereKeyStats(pParse, p, pRec, 0, a); 1559 iLower = a[0]; 1560 iUpper = a[0] + a[1]; 1561 } 1562 1563 assert( pLower==0 || (pLower->eOperator & (WO_GT|WO_GE))!=0 ); 1564 assert( pUpper==0 || (pUpper->eOperator & (WO_LT|WO_LE))!=0 ); 1565 assert( p->aSortOrder!=0 ); 1566 if( p->aSortOrder[nEq] ){ 1567 /* The roles of pLower and pUpper are swapped for a DESC index */ 1568 SWAP(WhereTerm*, pLower, pUpper); 1569 SWAP(int, nBtm, nTop); 1570 } 1571 1572 /* If possible, improve on the iLower estimate using ($P:$L). */ 1573 if( pLower ){ 1574 int n; /* Values extracted from pExpr */ 1575 Expr *pExpr = pLower->pExpr->pRight; 1576 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nBtm, nEq, &n); 1577 if( rc==SQLITE_OK && n ){ 1578 tRowcnt iNew; 1579 u16 mask = WO_GT|WO_LE; 1580 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1581 iLwrIdx = whereKeyStats(pParse, p, pRec, 0, a); 1582 iNew = a[0] + ((pLower->eOperator & mask) ? a[1] : 0); 1583 if( iNew>iLower ) iLower = iNew; 1584 nOut--; 1585 pLower = 0; 1586 } 1587 } 1588 1589 /* If possible, improve on the iUpper estimate using ($P:$U). */ 1590 if( pUpper ){ 1591 int n; /* Values extracted from pExpr */ 1592 Expr *pExpr = pUpper->pExpr->pRight; 1593 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, nTop, nEq, &n); 1594 if( rc==SQLITE_OK && n ){ 1595 tRowcnt iNew; 1596 u16 mask = WO_GT|WO_LE; 1597 if( sqlite3ExprVectorSize(pExpr)>n ) mask = (WO_LE|WO_LT); 1598 iUprIdx = whereKeyStats(pParse, p, pRec, 1, a); 1599 iNew = a[0] + ((pUpper->eOperator & mask) ? a[1] : 0); 1600 if( iNew<iUpper ) iUpper = iNew; 1601 nOut--; 1602 pUpper = 0; 1603 } 1604 } 1605 1606 pBuilder->pRec = pRec; 1607 if( rc==SQLITE_OK ){ 1608 if( iUpper>iLower ){ 1609 nNew = sqlite3LogEst(iUpper - iLower); 1610 /* TUNING: If both iUpper and iLower are derived from the same 1611 ** sample, then assume they are 4x more selective. This brings 1612 ** the estimated selectivity more in line with what it would be 1613 ** if estimated without the use of STAT4 tables. */ 1614 if( iLwrIdx==iUprIdx ) nNew -= 20; assert( 20==sqlite3LogEst(4) ); 1615 }else{ 1616 nNew = 10; assert( 10==sqlite3LogEst(2) ); 1617 } 1618 if( nNew<nOut ){ 1619 nOut = nNew; 1620 } 1621 WHERETRACE(0x10, ("STAT4 range scan: %u..%u est=%d\n", 1622 (u32)iLower, (u32)iUpper, nOut)); 1623 } 1624 }else{ 1625 int bDone = 0; 1626 rc = whereRangeSkipScanEst(pParse, pLower, pUpper, pLoop, &bDone); 1627 if( bDone ) return rc; 1628 } 1629 } 1630 #else 1631 UNUSED_PARAMETER(pParse); 1632 UNUSED_PARAMETER(pBuilder); 1633 assert( pLower || pUpper ); 1634 #endif 1635 assert( pUpper==0 || (pUpper->wtFlags & TERM_VNULL)==0 ); 1636 nNew = whereRangeAdjust(pLower, nOut); 1637 nNew = whereRangeAdjust(pUpper, nNew); 1638 1639 /* TUNING: If there is both an upper and lower limit and neither limit 1640 ** has an application-defined likelihood(), assume the range is 1641 ** reduced by an additional 75%. This means that, by default, an open-ended 1642 ** range query (e.g. col > ?) is assumed to match 1/4 of the rows in the 1643 ** index. While a closed range (e.g. col BETWEEN ? AND ?) is estimated to 1644 ** match 1/64 of the index. */ 1645 if( pLower && pLower->truthProb>0 && pUpper && pUpper->truthProb>0 ){ 1646 nNew -= 20; 1647 } 1648 1649 nOut -= (pLower!=0) + (pUpper!=0); 1650 if( nNew<10 ) nNew = 10; 1651 if( nNew<nOut ) nOut = nNew; 1652 #if defined(WHERETRACE_ENABLED) 1653 if( pLoop->nOut>nOut ){ 1654 WHERETRACE(0x10,("Range scan lowers nOut from %d to %d\n", 1655 pLoop->nOut, nOut)); 1656 } 1657 #endif 1658 pLoop->nOut = (LogEst)nOut; 1659 return rc; 1660 } 1661 1662 #ifdef SQLITE_ENABLE_STAT4 1663 /* 1664 ** Estimate the number of rows that will be returned based on 1665 ** an equality constraint x=VALUE and where that VALUE occurs in 1666 ** the histogram data. This only works when x is the left-most 1667 ** column of an index and sqlite_stat4 histogram data is available 1668 ** for that index. When pExpr==NULL that means the constraint is 1669 ** "x IS NULL" instead of "x=VALUE". 1670 ** 1671 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1672 ** If unable to make an estimate, leave *pnRow unchanged and return 1673 ** non-zero. 1674 ** 1675 ** This routine can fail if it is unable to load a collating sequence 1676 ** required for string comparison, or if unable to allocate memory 1677 ** for a UTF conversion required for comparison. The error is stored 1678 ** in the pParse structure. 1679 */ 1680 static int whereEqualScanEst( 1681 Parse *pParse, /* Parsing & code generating context */ 1682 WhereLoopBuilder *pBuilder, 1683 Expr *pExpr, /* Expression for VALUE in the x=VALUE constraint */ 1684 tRowcnt *pnRow /* Write the revised row estimate here */ 1685 ){ 1686 Index *p = pBuilder->pNew->u.btree.pIndex; 1687 int nEq = pBuilder->pNew->u.btree.nEq; 1688 UnpackedRecord *pRec = pBuilder->pRec; 1689 int rc; /* Subfunction return code */ 1690 tRowcnt a[2]; /* Statistics */ 1691 int bOk; 1692 1693 assert( nEq>=1 ); 1694 assert( nEq<=p->nColumn ); 1695 assert( p->aSample!=0 ); 1696 assert( p->nSample>0 ); 1697 assert( pBuilder->nRecValid<nEq ); 1698 1699 /* If values are not available for all fields of the index to the left 1700 ** of this one, no estimate can be made. Return SQLITE_NOTFOUND. */ 1701 if( pBuilder->nRecValid<(nEq-1) ){ 1702 return SQLITE_NOTFOUND; 1703 } 1704 1705 /* This is an optimization only. The call to sqlite3Stat4ProbeSetValue() 1706 ** below would return the same value. */ 1707 if( nEq>=p->nColumn ){ 1708 *pnRow = 1; 1709 return SQLITE_OK; 1710 } 1711 1712 rc = sqlite3Stat4ProbeSetValue(pParse, p, &pRec, pExpr, 1, nEq-1, &bOk); 1713 pBuilder->pRec = pRec; 1714 if( rc!=SQLITE_OK ) return rc; 1715 if( bOk==0 ) return SQLITE_NOTFOUND; 1716 pBuilder->nRecValid = nEq; 1717 1718 whereKeyStats(pParse, p, pRec, 0, a); 1719 WHERETRACE(0x10,("equality scan regions %s(%d): %d\n", 1720 p->zName, nEq-1, (int)a[1])); 1721 *pnRow = a[1]; 1722 1723 return rc; 1724 } 1725 #endif /* SQLITE_ENABLE_STAT4 */ 1726 1727 #ifdef SQLITE_ENABLE_STAT4 1728 /* 1729 ** Estimate the number of rows that will be returned based on 1730 ** an IN constraint where the right-hand side of the IN operator 1731 ** is a list of values. Example: 1732 ** 1733 ** WHERE x IN (1,2,3,4) 1734 ** 1735 ** Write the estimated row count into *pnRow and return SQLITE_OK. 1736 ** If unable to make an estimate, leave *pnRow unchanged and return 1737 ** non-zero. 1738 ** 1739 ** This routine can fail if it is unable to load a collating sequence 1740 ** required for string comparison, or if unable to allocate memory 1741 ** for a UTF conversion required for comparison. The error is stored 1742 ** in the pParse structure. 1743 */ 1744 static int whereInScanEst( 1745 Parse *pParse, /* Parsing & code generating context */ 1746 WhereLoopBuilder *pBuilder, 1747 ExprList *pList, /* The value list on the RHS of "x IN (v1,v2,v3,...)" */ 1748 tRowcnt *pnRow /* Write the revised row estimate here */ 1749 ){ 1750 Index *p = pBuilder->pNew->u.btree.pIndex; 1751 i64 nRow0 = sqlite3LogEstToInt(p->aiRowLogEst[0]); 1752 int nRecValid = pBuilder->nRecValid; 1753 int rc = SQLITE_OK; /* Subfunction return code */ 1754 tRowcnt nEst; /* Number of rows for a single term */ 1755 tRowcnt nRowEst = 0; /* New estimate of the number of rows */ 1756 int i; /* Loop counter */ 1757 1758 assert( p->aSample!=0 ); 1759 for(i=0; rc==SQLITE_OK && i<pList->nExpr; i++){ 1760 nEst = nRow0; 1761 rc = whereEqualScanEst(pParse, pBuilder, pList->a[i].pExpr, &nEst); 1762 nRowEst += nEst; 1763 pBuilder->nRecValid = nRecValid; 1764 } 1765 1766 if( rc==SQLITE_OK ){ 1767 if( nRowEst > nRow0 ) nRowEst = nRow0; 1768 *pnRow = nRowEst; 1769 WHERETRACE(0x10,("IN row estimate: est=%d\n", nRowEst)); 1770 } 1771 assert( pBuilder->nRecValid==nRecValid ); 1772 return rc; 1773 } 1774 #endif /* SQLITE_ENABLE_STAT4 */ 1775 1776 1777 #ifdef WHERETRACE_ENABLED 1778 /* 1779 ** Print the content of a WhereTerm object 1780 */ 1781 void sqlite3WhereTermPrint(WhereTerm *pTerm, int iTerm){ 1782 if( pTerm==0 ){ 1783 sqlite3DebugPrintf("TERM-%-3d NULL\n", iTerm); 1784 }else{ 1785 char zType[8]; 1786 char zLeft[50]; 1787 memcpy(zType, "....", 5); 1788 if( pTerm->wtFlags & TERM_VIRTUAL ) zType[0] = 'V'; 1789 if( pTerm->eOperator & WO_EQUIV ) zType[1] = 'E'; 1790 if( ExprHasProperty(pTerm->pExpr, EP_FromJoin) ) zType[2] = 'L'; 1791 if( pTerm->wtFlags & TERM_CODED ) zType[3] = 'C'; 1792 if( pTerm->eOperator & WO_SINGLE ){ 1793 sqlite3_snprintf(sizeof(zLeft),zLeft,"left={%d:%d}", 1794 pTerm->leftCursor, pTerm->u.x.leftColumn); 1795 }else if( (pTerm->eOperator & WO_OR)!=0 && pTerm->u.pOrInfo!=0 ){ 1796 sqlite3_snprintf(sizeof(zLeft),zLeft,"indexable=0x%llx", 1797 pTerm->u.pOrInfo->indexable); 1798 }else{ 1799 sqlite3_snprintf(sizeof(zLeft),zLeft,"left=%d", pTerm->leftCursor); 1800 } 1801 sqlite3DebugPrintf( 1802 "TERM-%-3d %p %s %-12s op=%03x wtFlags=%04x", 1803 iTerm, pTerm, zType, zLeft, pTerm->eOperator, pTerm->wtFlags); 1804 /* The 0x10000 .wheretrace flag causes extra information to be 1805 ** shown about each Term */ 1806 if( sqlite3WhereTrace & 0x10000 ){ 1807 sqlite3DebugPrintf(" prob=%-3d prereq=%llx,%llx", 1808 pTerm->truthProb, (u64)pTerm->prereqAll, (u64)pTerm->prereqRight); 1809 } 1810 if( pTerm->u.x.iField ){ 1811 sqlite3DebugPrintf(" iField=%d", pTerm->u.x.iField); 1812 } 1813 if( pTerm->iParent>=0 ){ 1814 sqlite3DebugPrintf(" iParent=%d", pTerm->iParent); 1815 } 1816 sqlite3DebugPrintf("\n"); 1817 sqlite3TreeViewExpr(0, pTerm->pExpr, 0); 1818 } 1819 } 1820 #endif 1821 1822 #ifdef WHERETRACE_ENABLED 1823 /* 1824 ** Show the complete content of a WhereClause 1825 */ 1826 void sqlite3WhereClausePrint(WhereClause *pWC){ 1827 int i; 1828 for(i=0; i<pWC->nTerm; i++){ 1829 sqlite3WhereTermPrint(&pWC->a[i], i); 1830 } 1831 } 1832 #endif 1833 1834 #ifdef WHERETRACE_ENABLED 1835 /* 1836 ** Print a WhereLoop object for debugging purposes 1837 */ 1838 void sqlite3WhereLoopPrint(WhereLoop *p, WhereClause *pWC){ 1839 WhereInfo *pWInfo = pWC->pWInfo; 1840 int nb = 1+(pWInfo->pTabList->nSrc+3)/4; 1841 SrcItem *pItem = pWInfo->pTabList->a + p->iTab; 1842 Table *pTab = pItem->pTab; 1843 Bitmask mAll = (((Bitmask)1)<<(nb*4)) - 1; 1844 sqlite3DebugPrintf("%c%2d.%0*llx.%0*llx", p->cId, 1845 p->iTab, nb, p->maskSelf, nb, p->prereq & mAll); 1846 sqlite3DebugPrintf(" %12s", 1847 pItem->zAlias ? pItem->zAlias : pTab->zName); 1848 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 1849 const char *zName; 1850 if( p->u.btree.pIndex && (zName = p->u.btree.pIndex->zName)!=0 ){ 1851 if( strncmp(zName, "sqlite_autoindex_", 17)==0 ){ 1852 int i = sqlite3Strlen30(zName) - 1; 1853 while( zName[i]!='_' ) i--; 1854 zName += i; 1855 } 1856 sqlite3DebugPrintf(".%-16s %2d", zName, p->u.btree.nEq); 1857 }else{ 1858 sqlite3DebugPrintf("%20s",""); 1859 } 1860 }else{ 1861 char *z; 1862 if( p->u.vtab.idxStr ){ 1863 z = sqlite3_mprintf("(%d,\"%s\",%#x)", 1864 p->u.vtab.idxNum, p->u.vtab.idxStr, p->u.vtab.omitMask); 1865 }else{ 1866 z = sqlite3_mprintf("(%d,%x)", p->u.vtab.idxNum, p->u.vtab.omitMask); 1867 } 1868 sqlite3DebugPrintf(" %-19s", z); 1869 sqlite3_free(z); 1870 } 1871 if( p->wsFlags & WHERE_SKIPSCAN ){ 1872 sqlite3DebugPrintf(" f %05x %d-%d", p->wsFlags, p->nLTerm,p->nSkip); 1873 }else{ 1874 sqlite3DebugPrintf(" f %05x N %d", p->wsFlags, p->nLTerm); 1875 } 1876 sqlite3DebugPrintf(" cost %d,%d,%d\n", p->rSetup, p->rRun, p->nOut); 1877 if( p->nLTerm && (sqlite3WhereTrace & 0x100)!=0 ){ 1878 int i; 1879 for(i=0; i<p->nLTerm; i++){ 1880 sqlite3WhereTermPrint(p->aLTerm[i], i); 1881 } 1882 } 1883 } 1884 #endif 1885 1886 /* 1887 ** Convert bulk memory into a valid WhereLoop that can be passed 1888 ** to whereLoopClear harmlessly. 1889 */ 1890 static void whereLoopInit(WhereLoop *p){ 1891 p->aLTerm = p->aLTermSpace; 1892 p->nLTerm = 0; 1893 p->nLSlot = ArraySize(p->aLTermSpace); 1894 p->wsFlags = 0; 1895 } 1896 1897 /* 1898 ** Clear the WhereLoop.u union. Leave WhereLoop.pLTerm intact. 1899 */ 1900 static void whereLoopClearUnion(sqlite3 *db, WhereLoop *p){ 1901 if( p->wsFlags & (WHERE_VIRTUALTABLE|WHERE_AUTO_INDEX) ){ 1902 if( (p->wsFlags & WHERE_VIRTUALTABLE)!=0 && p->u.vtab.needFree ){ 1903 sqlite3_free(p->u.vtab.idxStr); 1904 p->u.vtab.needFree = 0; 1905 p->u.vtab.idxStr = 0; 1906 }else if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 && p->u.btree.pIndex!=0 ){ 1907 sqlite3DbFree(db, p->u.btree.pIndex->zColAff); 1908 sqlite3DbFreeNN(db, p->u.btree.pIndex); 1909 p->u.btree.pIndex = 0; 1910 } 1911 } 1912 } 1913 1914 /* 1915 ** Deallocate internal memory used by a WhereLoop object 1916 */ 1917 static void whereLoopClear(sqlite3 *db, WhereLoop *p){ 1918 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1919 whereLoopClearUnion(db, p); 1920 whereLoopInit(p); 1921 } 1922 1923 /* 1924 ** Increase the memory allocation for pLoop->aLTerm[] to be at least n. 1925 */ 1926 static int whereLoopResize(sqlite3 *db, WhereLoop *p, int n){ 1927 WhereTerm **paNew; 1928 if( p->nLSlot>=n ) return SQLITE_OK; 1929 n = (n+7)&~7; 1930 paNew = sqlite3DbMallocRawNN(db, sizeof(p->aLTerm[0])*n); 1931 if( paNew==0 ) return SQLITE_NOMEM_BKPT; 1932 memcpy(paNew, p->aLTerm, sizeof(p->aLTerm[0])*p->nLSlot); 1933 if( p->aLTerm!=p->aLTermSpace ) sqlite3DbFreeNN(db, p->aLTerm); 1934 p->aLTerm = paNew; 1935 p->nLSlot = n; 1936 return SQLITE_OK; 1937 } 1938 1939 /* 1940 ** Transfer content from the second pLoop into the first. 1941 */ 1942 static int whereLoopXfer(sqlite3 *db, WhereLoop *pTo, WhereLoop *pFrom){ 1943 whereLoopClearUnion(db, pTo); 1944 if( whereLoopResize(db, pTo, pFrom->nLTerm) ){ 1945 memset(pTo, 0, WHERE_LOOP_XFER_SZ); 1946 return SQLITE_NOMEM_BKPT; 1947 } 1948 memcpy(pTo, pFrom, WHERE_LOOP_XFER_SZ); 1949 memcpy(pTo->aLTerm, pFrom->aLTerm, pTo->nLTerm*sizeof(pTo->aLTerm[0])); 1950 if( pFrom->wsFlags & WHERE_VIRTUALTABLE ){ 1951 pFrom->u.vtab.needFree = 0; 1952 }else if( (pFrom->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 1953 pFrom->u.btree.pIndex = 0; 1954 } 1955 return SQLITE_OK; 1956 } 1957 1958 /* 1959 ** Delete a WhereLoop object 1960 */ 1961 static void whereLoopDelete(sqlite3 *db, WhereLoop *p){ 1962 whereLoopClear(db, p); 1963 sqlite3DbFreeNN(db, p); 1964 } 1965 1966 /* 1967 ** Free a WhereInfo structure 1968 */ 1969 static void whereInfoFree(sqlite3 *db, WhereInfo *pWInfo){ 1970 int i; 1971 assert( pWInfo!=0 ); 1972 for(i=0; i<pWInfo->nLevel; i++){ 1973 WhereLevel *pLevel = &pWInfo->a[i]; 1974 if( pLevel->pWLoop && (pLevel->pWLoop->wsFlags & WHERE_IN_ABLE) ){ 1975 sqlite3DbFree(db, pLevel->u.in.aInLoop); 1976 } 1977 } 1978 sqlite3WhereClauseClear(&pWInfo->sWC); 1979 while( pWInfo->pLoops ){ 1980 WhereLoop *p = pWInfo->pLoops; 1981 pWInfo->pLoops = p->pNextLoop; 1982 whereLoopDelete(db, p); 1983 } 1984 assert( pWInfo->pExprMods==0 ); 1985 sqlite3DbFreeNN(db, pWInfo); 1986 } 1987 1988 /* Undo all Expr node modifications 1989 */ 1990 static void whereUndoExprMods(WhereInfo *pWInfo){ 1991 while( pWInfo->pExprMods ){ 1992 WhereExprMod *p = pWInfo->pExprMods; 1993 pWInfo->pExprMods = p->pNext; 1994 memcpy(p->pExpr, &p->orig, sizeof(p->orig)); 1995 sqlite3DbFree(pWInfo->pParse->db, p); 1996 } 1997 } 1998 1999 /* 2000 ** Return TRUE if all of the following are true: 2001 ** 2002 ** (1) X has the same or lower cost that Y 2003 ** (2) X uses fewer WHERE clause terms than Y 2004 ** (3) Every WHERE clause term used by X is also used by Y 2005 ** (4) X skips at least as many columns as Y 2006 ** (5) If X is a covering index, than Y is too 2007 ** 2008 ** Conditions (2) and (3) mean that X is a "proper subset" of Y. 2009 ** If X is a proper subset of Y then Y is a better choice and ought 2010 ** to have a lower cost. This routine returns TRUE when that cost 2011 ** relationship is inverted and needs to be adjusted. Constraint (4) 2012 ** was added because if X uses skip-scan less than Y it still might 2013 ** deserve a lower cost even if it is a proper subset of Y. Constraint (5) 2014 ** was added because a covering index probably deserves to have a lower cost 2015 ** than a non-covering index even if it is a proper subset. 2016 */ 2017 static int whereLoopCheaperProperSubset( 2018 const WhereLoop *pX, /* First WhereLoop to compare */ 2019 const WhereLoop *pY /* Compare against this WhereLoop */ 2020 ){ 2021 int i, j; 2022 if( pX->nLTerm-pX->nSkip >= pY->nLTerm-pY->nSkip ){ 2023 return 0; /* X is not a subset of Y */ 2024 } 2025 if( pY->nSkip > pX->nSkip ) return 0; 2026 if( pX->rRun >= pY->rRun ){ 2027 if( pX->rRun > pY->rRun ) return 0; /* X costs more than Y */ 2028 if( pX->nOut > pY->nOut ) return 0; /* X costs more than Y */ 2029 } 2030 for(i=pX->nLTerm-1; i>=0; i--){ 2031 if( pX->aLTerm[i]==0 ) continue; 2032 for(j=pY->nLTerm-1; j>=0; j--){ 2033 if( pY->aLTerm[j]==pX->aLTerm[i] ) break; 2034 } 2035 if( j<0 ) return 0; /* X not a subset of Y since term X[i] not used by Y */ 2036 } 2037 if( (pX->wsFlags&WHERE_IDX_ONLY)!=0 2038 && (pY->wsFlags&WHERE_IDX_ONLY)==0 ){ 2039 return 0; /* Constraint (5) */ 2040 } 2041 return 1; /* All conditions meet */ 2042 } 2043 2044 /* 2045 ** Try to adjust the cost of WhereLoop pTemplate upwards or downwards so 2046 ** that: 2047 ** 2048 ** (1) pTemplate costs less than any other WhereLoops that are a proper 2049 ** subset of pTemplate 2050 ** 2051 ** (2) pTemplate costs more than any other WhereLoops for which pTemplate 2052 ** is a proper subset. 2053 ** 2054 ** To say "WhereLoop X is a proper subset of Y" means that X uses fewer 2055 ** WHERE clause terms than Y and that every WHERE clause term used by X is 2056 ** also used by Y. 2057 */ 2058 static void whereLoopAdjustCost(const WhereLoop *p, WhereLoop *pTemplate){ 2059 if( (pTemplate->wsFlags & WHERE_INDEXED)==0 ) return; 2060 for(; p; p=p->pNextLoop){ 2061 if( p->iTab!=pTemplate->iTab ) continue; 2062 if( (p->wsFlags & WHERE_INDEXED)==0 ) continue; 2063 if( whereLoopCheaperProperSubset(p, pTemplate) ){ 2064 /* Adjust pTemplate cost downward so that it is cheaper than its 2065 ** subset p. */ 2066 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2067 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut-1)); 2068 pTemplate->rRun = p->rRun; 2069 pTemplate->nOut = p->nOut - 1; 2070 }else if( whereLoopCheaperProperSubset(pTemplate, p) ){ 2071 /* Adjust pTemplate cost upward so that it is costlier than p since 2072 ** pTemplate is a proper subset of p */ 2073 WHERETRACE(0x80,("subset cost adjustment %d,%d to %d,%d\n", 2074 pTemplate->rRun, pTemplate->nOut, p->rRun, p->nOut+1)); 2075 pTemplate->rRun = p->rRun; 2076 pTemplate->nOut = p->nOut + 1; 2077 } 2078 } 2079 } 2080 2081 /* 2082 ** Search the list of WhereLoops in *ppPrev looking for one that can be 2083 ** replaced by pTemplate. 2084 ** 2085 ** Return NULL if pTemplate does not belong on the WhereLoop list. 2086 ** In other words if pTemplate ought to be dropped from further consideration. 2087 ** 2088 ** If pX is a WhereLoop that pTemplate can replace, then return the 2089 ** link that points to pX. 2090 ** 2091 ** If pTemplate cannot replace any existing element of the list but needs 2092 ** to be added to the list as a new entry, then return a pointer to the 2093 ** tail of the list. 2094 */ 2095 static WhereLoop **whereLoopFindLesser( 2096 WhereLoop **ppPrev, 2097 const WhereLoop *pTemplate 2098 ){ 2099 WhereLoop *p; 2100 for(p=(*ppPrev); p; ppPrev=&p->pNextLoop, p=*ppPrev){ 2101 if( p->iTab!=pTemplate->iTab || p->iSortIdx!=pTemplate->iSortIdx ){ 2102 /* If either the iTab or iSortIdx values for two WhereLoop are different 2103 ** then those WhereLoops need to be considered separately. Neither is 2104 ** a candidate to replace the other. */ 2105 continue; 2106 } 2107 /* In the current implementation, the rSetup value is either zero 2108 ** or the cost of building an automatic index (NlogN) and the NlogN 2109 ** is the same for compatible WhereLoops. */ 2110 assert( p->rSetup==0 || pTemplate->rSetup==0 2111 || p->rSetup==pTemplate->rSetup ); 2112 2113 /* whereLoopAddBtree() always generates and inserts the automatic index 2114 ** case first. Hence compatible candidate WhereLoops never have a larger 2115 ** rSetup. Call this SETUP-INVARIANT */ 2116 assert( p->rSetup>=pTemplate->rSetup ); 2117 2118 /* Any loop using an appliation-defined index (or PRIMARY KEY or 2119 ** UNIQUE constraint) with one or more == constraints is better 2120 ** than an automatic index. Unless it is a skip-scan. */ 2121 if( (p->wsFlags & WHERE_AUTO_INDEX)!=0 2122 && (pTemplate->nSkip)==0 2123 && (pTemplate->wsFlags & WHERE_INDEXED)!=0 2124 && (pTemplate->wsFlags & WHERE_COLUMN_EQ)!=0 2125 && (p->prereq & pTemplate->prereq)==pTemplate->prereq 2126 ){ 2127 break; 2128 } 2129 2130 /* If existing WhereLoop p is better than pTemplate, pTemplate can be 2131 ** discarded. WhereLoop p is better if: 2132 ** (1) p has no more dependencies than pTemplate, and 2133 ** (2) p has an equal or lower cost than pTemplate 2134 */ 2135 if( (p->prereq & pTemplate->prereq)==p->prereq /* (1) */ 2136 && p->rSetup<=pTemplate->rSetup /* (2a) */ 2137 && p->rRun<=pTemplate->rRun /* (2b) */ 2138 && p->nOut<=pTemplate->nOut /* (2c) */ 2139 ){ 2140 return 0; /* Discard pTemplate */ 2141 } 2142 2143 /* If pTemplate is always better than p, then cause p to be overwritten 2144 ** with pTemplate. pTemplate is better than p if: 2145 ** (1) pTemplate has no more dependences than p, and 2146 ** (2) pTemplate has an equal or lower cost than p. 2147 */ 2148 if( (p->prereq & pTemplate->prereq)==pTemplate->prereq /* (1) */ 2149 && p->rRun>=pTemplate->rRun /* (2a) */ 2150 && p->nOut>=pTemplate->nOut /* (2b) */ 2151 ){ 2152 assert( p->rSetup>=pTemplate->rSetup ); /* SETUP-INVARIANT above */ 2153 break; /* Cause p to be overwritten by pTemplate */ 2154 } 2155 } 2156 return ppPrev; 2157 } 2158 2159 /* 2160 ** Insert or replace a WhereLoop entry using the template supplied. 2161 ** 2162 ** An existing WhereLoop entry might be overwritten if the new template 2163 ** is better and has fewer dependencies. Or the template will be ignored 2164 ** and no insert will occur if an existing WhereLoop is faster and has 2165 ** fewer dependencies than the template. Otherwise a new WhereLoop is 2166 ** added based on the template. 2167 ** 2168 ** If pBuilder->pOrSet is not NULL then we care about only the 2169 ** prerequisites and rRun and nOut costs of the N best loops. That 2170 ** information is gathered in the pBuilder->pOrSet object. This special 2171 ** processing mode is used only for OR clause processing. 2172 ** 2173 ** When accumulating multiple loops (when pBuilder->pOrSet is NULL) we 2174 ** still might overwrite similar loops with the new template if the 2175 ** new template is better. Loops may be overwritten if the following 2176 ** conditions are met: 2177 ** 2178 ** (1) They have the same iTab. 2179 ** (2) They have the same iSortIdx. 2180 ** (3) The template has same or fewer dependencies than the current loop 2181 ** (4) The template has the same or lower cost than the current loop 2182 */ 2183 static int whereLoopInsert(WhereLoopBuilder *pBuilder, WhereLoop *pTemplate){ 2184 WhereLoop **ppPrev, *p; 2185 WhereInfo *pWInfo = pBuilder->pWInfo; 2186 sqlite3 *db = pWInfo->pParse->db; 2187 int rc; 2188 2189 /* Stop the search once we hit the query planner search limit */ 2190 if( pBuilder->iPlanLimit==0 ){ 2191 WHERETRACE(0xffffffff,("=== query planner search limit reached ===\n")); 2192 if( pBuilder->pOrSet ) pBuilder->pOrSet->n = 0; 2193 return SQLITE_DONE; 2194 } 2195 pBuilder->iPlanLimit--; 2196 2197 whereLoopAdjustCost(pWInfo->pLoops, pTemplate); 2198 2199 /* If pBuilder->pOrSet is defined, then only keep track of the costs 2200 ** and prereqs. 2201 */ 2202 if( pBuilder->pOrSet!=0 ){ 2203 if( pTemplate->nLTerm ){ 2204 #if WHERETRACE_ENABLED 2205 u16 n = pBuilder->pOrSet->n; 2206 int x = 2207 #endif 2208 whereOrInsert(pBuilder->pOrSet, pTemplate->prereq, pTemplate->rRun, 2209 pTemplate->nOut); 2210 #if WHERETRACE_ENABLED /* 0x8 */ 2211 if( sqlite3WhereTrace & 0x8 ){ 2212 sqlite3DebugPrintf(x?" or-%d: ":" or-X: ", n); 2213 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2214 } 2215 #endif 2216 } 2217 return SQLITE_OK; 2218 } 2219 2220 /* Look for an existing WhereLoop to replace with pTemplate 2221 */ 2222 ppPrev = whereLoopFindLesser(&pWInfo->pLoops, pTemplate); 2223 2224 if( ppPrev==0 ){ 2225 /* There already exists a WhereLoop on the list that is better 2226 ** than pTemplate, so just ignore pTemplate */ 2227 #if WHERETRACE_ENABLED /* 0x8 */ 2228 if( sqlite3WhereTrace & 0x8 ){ 2229 sqlite3DebugPrintf(" skip: "); 2230 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2231 } 2232 #endif 2233 return SQLITE_OK; 2234 }else{ 2235 p = *ppPrev; 2236 } 2237 2238 /* If we reach this point it means that either p[] should be overwritten 2239 ** with pTemplate[] if p[] exists, or if p==NULL then allocate a new 2240 ** WhereLoop and insert it. 2241 */ 2242 #if WHERETRACE_ENABLED /* 0x8 */ 2243 if( sqlite3WhereTrace & 0x8 ){ 2244 if( p!=0 ){ 2245 sqlite3DebugPrintf("replace: "); 2246 sqlite3WhereLoopPrint(p, pBuilder->pWC); 2247 sqlite3DebugPrintf(" with: "); 2248 }else{ 2249 sqlite3DebugPrintf(" add: "); 2250 } 2251 sqlite3WhereLoopPrint(pTemplate, pBuilder->pWC); 2252 } 2253 #endif 2254 if( p==0 ){ 2255 /* Allocate a new WhereLoop to add to the end of the list */ 2256 *ppPrev = p = sqlite3DbMallocRawNN(db, sizeof(WhereLoop)); 2257 if( p==0 ) return SQLITE_NOMEM_BKPT; 2258 whereLoopInit(p); 2259 p->pNextLoop = 0; 2260 }else{ 2261 /* We will be overwriting WhereLoop p[]. But before we do, first 2262 ** go through the rest of the list and delete any other entries besides 2263 ** p[] that are also supplated by pTemplate */ 2264 WhereLoop **ppTail = &p->pNextLoop; 2265 WhereLoop *pToDel; 2266 while( *ppTail ){ 2267 ppTail = whereLoopFindLesser(ppTail, pTemplate); 2268 if( ppTail==0 ) break; 2269 pToDel = *ppTail; 2270 if( pToDel==0 ) break; 2271 *ppTail = pToDel->pNextLoop; 2272 #if WHERETRACE_ENABLED /* 0x8 */ 2273 if( sqlite3WhereTrace & 0x8 ){ 2274 sqlite3DebugPrintf(" delete: "); 2275 sqlite3WhereLoopPrint(pToDel, pBuilder->pWC); 2276 } 2277 #endif 2278 whereLoopDelete(db, pToDel); 2279 } 2280 } 2281 rc = whereLoopXfer(db, p, pTemplate); 2282 if( (p->wsFlags & WHERE_VIRTUALTABLE)==0 ){ 2283 Index *pIndex = p->u.btree.pIndex; 2284 if( pIndex && pIndex->idxType==SQLITE_IDXTYPE_IPK ){ 2285 p->u.btree.pIndex = 0; 2286 } 2287 } 2288 return rc; 2289 } 2290 2291 /* 2292 ** Adjust the WhereLoop.nOut value downward to account for terms of the 2293 ** WHERE clause that reference the loop but which are not used by an 2294 ** index. 2295 * 2296 ** For every WHERE clause term that is not used by the index 2297 ** and which has a truth probability assigned by one of the likelihood(), 2298 ** likely(), or unlikely() SQL functions, reduce the estimated number 2299 ** of output rows by the probability specified. 2300 ** 2301 ** TUNING: For every WHERE clause term that is not used by the index 2302 ** and which does not have an assigned truth probability, heuristics 2303 ** described below are used to try to estimate the truth probability. 2304 ** TODO --> Perhaps this is something that could be improved by better 2305 ** table statistics. 2306 ** 2307 ** Heuristic 1: Estimate the truth probability as 93.75%. The 93.75% 2308 ** value corresponds to -1 in LogEst notation, so this means decrement 2309 ** the WhereLoop.nOut field for every such WHERE clause term. 2310 ** 2311 ** Heuristic 2: If there exists one or more WHERE clause terms of the 2312 ** form "x==EXPR" and EXPR is not a constant 0 or 1, then make sure the 2313 ** final output row estimate is no greater than 1/4 of the total number 2314 ** of rows in the table. In other words, assume that x==EXPR will filter 2315 ** out at least 3 out of 4 rows. If EXPR is -1 or 0 or 1, then maybe the 2316 ** "x" column is boolean or else -1 or 0 or 1 is a common default value 2317 ** on the "x" column and so in that case only cap the output row estimate 2318 ** at 1/2 instead of 1/4. 2319 */ 2320 static void whereLoopOutputAdjust( 2321 WhereClause *pWC, /* The WHERE clause */ 2322 WhereLoop *pLoop, /* The loop to adjust downward */ 2323 LogEst nRow /* Number of rows in the entire table */ 2324 ){ 2325 WhereTerm *pTerm, *pX; 2326 Bitmask notAllowed = ~(pLoop->prereq|pLoop->maskSelf); 2327 int i, j; 2328 LogEst iReduce = 0; /* pLoop->nOut should not exceed nRow-iReduce */ 2329 2330 assert( (pLoop->wsFlags & WHERE_AUTO_INDEX)==0 ); 2331 for(i=pWC->nTerm, pTerm=pWC->a; i>0; i--, pTerm++){ 2332 assert( pTerm!=0 ); 2333 if( (pTerm->wtFlags & TERM_VIRTUAL)!=0 ) break; 2334 if( (pTerm->prereqAll & pLoop->maskSelf)==0 ) continue; 2335 if( (pTerm->prereqAll & notAllowed)!=0 ) continue; 2336 for(j=pLoop->nLTerm-1; j>=0; j--){ 2337 pX = pLoop->aLTerm[j]; 2338 if( pX==0 ) continue; 2339 if( pX==pTerm ) break; 2340 if( pX->iParent>=0 && (&pWC->a[pX->iParent])==pTerm ) break; 2341 } 2342 if( j<0 ){ 2343 if( pTerm->truthProb<=0 ){ 2344 /* If a truth probability is specified using the likelihood() hints, 2345 ** then use the probability provided by the application. */ 2346 pLoop->nOut += pTerm->truthProb; 2347 }else{ 2348 /* In the absence of explicit truth probabilities, use heuristics to 2349 ** guess a reasonable truth probability. */ 2350 pLoop->nOut--; 2351 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 2352 && (pTerm->wtFlags & TERM_HIGHTRUTH)==0 /* tag-20200224-1 */ 2353 ){ 2354 Expr *pRight = pTerm->pExpr->pRight; 2355 int k = 0; 2356 testcase( pTerm->pExpr->op==TK_IS ); 2357 if( sqlite3ExprIsInteger(pRight, &k) && k>=(-1) && k<=1 ){ 2358 k = 10; 2359 }else{ 2360 k = 20; 2361 } 2362 if( iReduce<k ){ 2363 pTerm->wtFlags |= TERM_HEURTRUTH; 2364 iReduce = k; 2365 } 2366 } 2367 } 2368 } 2369 } 2370 if( pLoop->nOut > nRow-iReduce ) pLoop->nOut = nRow - iReduce; 2371 } 2372 2373 /* 2374 ** Term pTerm is a vector range comparison operation. The first comparison 2375 ** in the vector can be optimized using column nEq of the index. This 2376 ** function returns the total number of vector elements that can be used 2377 ** as part of the range comparison. 2378 ** 2379 ** For example, if the query is: 2380 ** 2381 ** WHERE a = ? AND (b, c, d) > (?, ?, ?) 2382 ** 2383 ** and the index: 2384 ** 2385 ** CREATE INDEX ... ON (a, b, c, d, e) 2386 ** 2387 ** then this function would be invoked with nEq=1. The value returned in 2388 ** this case is 3. 2389 */ 2390 static int whereRangeVectorLen( 2391 Parse *pParse, /* Parsing context */ 2392 int iCur, /* Cursor open on pIdx */ 2393 Index *pIdx, /* The index to be used for a inequality constraint */ 2394 int nEq, /* Number of prior equality constraints on same index */ 2395 WhereTerm *pTerm /* The vector inequality constraint */ 2396 ){ 2397 int nCmp = sqlite3ExprVectorSize(pTerm->pExpr->pLeft); 2398 int i; 2399 2400 nCmp = MIN(nCmp, (pIdx->nColumn - nEq)); 2401 for(i=1; i<nCmp; i++){ 2402 /* Test if comparison i of pTerm is compatible with column (i+nEq) 2403 ** of the index. If not, exit the loop. */ 2404 char aff; /* Comparison affinity */ 2405 char idxaff = 0; /* Indexed columns affinity */ 2406 CollSeq *pColl; /* Comparison collation sequence */ 2407 Expr *pLhs = pTerm->pExpr->pLeft->x.pList->a[i].pExpr; 2408 Expr *pRhs = pTerm->pExpr->pRight; 2409 if( pRhs->flags & EP_xIsSelect ){ 2410 pRhs = pRhs->x.pSelect->pEList->a[i].pExpr; 2411 }else{ 2412 pRhs = pRhs->x.pList->a[i].pExpr; 2413 } 2414 2415 /* Check that the LHS of the comparison is a column reference to 2416 ** the right column of the right source table. And that the sort 2417 ** order of the index column is the same as the sort order of the 2418 ** leftmost index column. */ 2419 if( pLhs->op!=TK_COLUMN 2420 || pLhs->iTable!=iCur 2421 || pLhs->iColumn!=pIdx->aiColumn[i+nEq] 2422 || pIdx->aSortOrder[i+nEq]!=pIdx->aSortOrder[nEq] 2423 ){ 2424 break; 2425 } 2426 2427 testcase( pLhs->iColumn==XN_ROWID ); 2428 aff = sqlite3CompareAffinity(pRhs, sqlite3ExprAffinity(pLhs)); 2429 idxaff = sqlite3TableColumnAffinity(pIdx->pTable, pLhs->iColumn); 2430 if( aff!=idxaff ) break; 2431 2432 pColl = sqlite3BinaryCompareCollSeq(pParse, pLhs, pRhs); 2433 if( pColl==0 ) break; 2434 if( sqlite3StrICmp(pColl->zName, pIdx->azColl[i+nEq]) ) break; 2435 } 2436 return i; 2437 } 2438 2439 /* 2440 ** Adjust the cost C by the costMult facter T. This only occurs if 2441 ** compiled with -DSQLITE_ENABLE_COSTMULT 2442 */ 2443 #ifdef SQLITE_ENABLE_COSTMULT 2444 # define ApplyCostMultiplier(C,T) C += T 2445 #else 2446 # define ApplyCostMultiplier(C,T) 2447 #endif 2448 2449 /* 2450 ** We have so far matched pBuilder->pNew->u.btree.nEq terms of the 2451 ** index pIndex. Try to match one more. 2452 ** 2453 ** When this function is called, pBuilder->pNew->nOut contains the 2454 ** number of rows expected to be visited by filtering using the nEq 2455 ** terms only. If it is modified, this value is restored before this 2456 ** function returns. 2457 ** 2458 ** If pProbe->idxType==SQLITE_IDXTYPE_IPK, that means pIndex is 2459 ** a fake index used for the INTEGER PRIMARY KEY. 2460 */ 2461 static int whereLoopAddBtreeIndex( 2462 WhereLoopBuilder *pBuilder, /* The WhereLoop factory */ 2463 SrcItem *pSrc, /* FROM clause term being analyzed */ 2464 Index *pProbe, /* An index on pSrc */ 2465 LogEst nInMul /* log(Number of iterations due to IN) */ 2466 ){ 2467 WhereInfo *pWInfo = pBuilder->pWInfo; /* WHERE analyse context */ 2468 Parse *pParse = pWInfo->pParse; /* Parsing context */ 2469 sqlite3 *db = pParse->db; /* Database connection malloc context */ 2470 WhereLoop *pNew; /* Template WhereLoop under construction */ 2471 WhereTerm *pTerm; /* A WhereTerm under consideration */ 2472 int opMask; /* Valid operators for constraints */ 2473 WhereScan scan; /* Iterator for WHERE terms */ 2474 Bitmask saved_prereq; /* Original value of pNew->prereq */ 2475 u16 saved_nLTerm; /* Original value of pNew->nLTerm */ 2476 u16 saved_nEq; /* Original value of pNew->u.btree.nEq */ 2477 u16 saved_nBtm; /* Original value of pNew->u.btree.nBtm */ 2478 u16 saved_nTop; /* Original value of pNew->u.btree.nTop */ 2479 u16 saved_nSkip; /* Original value of pNew->nSkip */ 2480 u32 saved_wsFlags; /* Original value of pNew->wsFlags */ 2481 LogEst saved_nOut; /* Original value of pNew->nOut */ 2482 int rc = SQLITE_OK; /* Return code */ 2483 LogEst rSize; /* Number of rows in the table */ 2484 LogEst rLogSize; /* Logarithm of table size */ 2485 WhereTerm *pTop = 0, *pBtm = 0; /* Top and bottom range constraints */ 2486 2487 pNew = pBuilder->pNew; 2488 if( db->mallocFailed ) return SQLITE_NOMEM_BKPT; 2489 WHERETRACE(0x800, ("BEGIN %s.addBtreeIdx(%s), nEq=%d, nSkip=%d, rRun=%d\n", 2490 pProbe->pTable->zName,pProbe->zName, 2491 pNew->u.btree.nEq, pNew->nSkip, pNew->rRun)); 2492 2493 assert( (pNew->wsFlags & WHERE_VIRTUALTABLE)==0 ); 2494 assert( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 ); 2495 if( pNew->wsFlags & WHERE_BTM_LIMIT ){ 2496 opMask = WO_LT|WO_LE; 2497 }else{ 2498 assert( pNew->u.btree.nBtm==0 ); 2499 opMask = WO_EQ|WO_IN|WO_GT|WO_GE|WO_LT|WO_LE|WO_ISNULL|WO_IS; 2500 } 2501 if( pProbe->bUnordered ) opMask &= ~(WO_GT|WO_GE|WO_LT|WO_LE); 2502 2503 assert( pNew->u.btree.nEq<pProbe->nColumn ); 2504 2505 saved_nEq = pNew->u.btree.nEq; 2506 saved_nBtm = pNew->u.btree.nBtm; 2507 saved_nTop = pNew->u.btree.nTop; 2508 saved_nSkip = pNew->nSkip; 2509 saved_nLTerm = pNew->nLTerm; 2510 saved_wsFlags = pNew->wsFlags; 2511 saved_prereq = pNew->prereq; 2512 saved_nOut = pNew->nOut; 2513 pTerm = whereScanInit(&scan, pBuilder->pWC, pSrc->iCursor, saved_nEq, 2514 opMask, pProbe); 2515 pNew->rSetup = 0; 2516 rSize = pProbe->aiRowLogEst[0]; 2517 rLogSize = estLog(rSize); 2518 for(; rc==SQLITE_OK && pTerm!=0; pTerm = whereScanNext(&scan)){ 2519 u16 eOp = pTerm->eOperator; /* Shorthand for pTerm->eOperator */ 2520 LogEst rCostIdx; 2521 LogEst nOutUnadjusted; /* nOut before IN() and WHERE adjustments */ 2522 int nIn = 0; 2523 #ifdef SQLITE_ENABLE_STAT4 2524 int nRecValid = pBuilder->nRecValid; 2525 #endif 2526 if( (eOp==WO_ISNULL || (pTerm->wtFlags&TERM_VNULL)!=0) 2527 && indexColumnNotNull(pProbe, saved_nEq) 2528 ){ 2529 continue; /* ignore IS [NOT] NULL constraints on NOT NULL columns */ 2530 } 2531 if( pTerm->prereqRight & pNew->maskSelf ) continue; 2532 2533 /* Do not allow the upper bound of a LIKE optimization range constraint 2534 ** to mix with a lower range bound from some other source */ 2535 if( pTerm->wtFlags & TERM_LIKEOPT && pTerm->eOperator==WO_LT ) continue; 2536 2537 /* tag-20191211-001: Do not allow constraints from the WHERE clause to 2538 ** be used by the right table of a LEFT JOIN. Only constraints in the 2539 ** ON clause are allowed. See tag-20191211-002 for the vtab equivalent. */ 2540 if( (pSrc->fg.jointype & JT_LEFT)!=0 2541 && !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 2542 ){ 2543 continue; 2544 } 2545 2546 if( IsUniqueIndex(pProbe) && saved_nEq==pProbe->nKeyCol-1 ){ 2547 pBuilder->bldFlags1 |= SQLITE_BLDF1_UNIQUE; 2548 }else{ 2549 pBuilder->bldFlags1 |= SQLITE_BLDF1_INDEXED; 2550 } 2551 pNew->wsFlags = saved_wsFlags; 2552 pNew->u.btree.nEq = saved_nEq; 2553 pNew->u.btree.nBtm = saved_nBtm; 2554 pNew->u.btree.nTop = saved_nTop; 2555 pNew->nLTerm = saved_nLTerm; 2556 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2557 pNew->aLTerm[pNew->nLTerm++] = pTerm; 2558 pNew->prereq = (saved_prereq | pTerm->prereqRight) & ~pNew->maskSelf; 2559 2560 assert( nInMul==0 2561 || (pNew->wsFlags & WHERE_COLUMN_NULL)!=0 2562 || (pNew->wsFlags & WHERE_COLUMN_IN)!=0 2563 || (pNew->wsFlags & WHERE_SKIPSCAN)!=0 2564 ); 2565 2566 if( eOp & WO_IN ){ 2567 Expr *pExpr = pTerm->pExpr; 2568 if( ExprHasProperty(pExpr, EP_xIsSelect) ){ 2569 /* "x IN (SELECT ...)": TUNING: the SELECT returns 25 rows */ 2570 int i; 2571 nIn = 46; assert( 46==sqlite3LogEst(25) ); 2572 2573 /* The expression may actually be of the form (x, y) IN (SELECT...). 2574 ** In this case there is a separate term for each of (x) and (y). 2575 ** However, the nIn multiplier should only be applied once, not once 2576 ** for each such term. The following loop checks that pTerm is the 2577 ** first such term in use, and sets nIn back to 0 if it is not. */ 2578 for(i=0; i<pNew->nLTerm-1; i++){ 2579 if( pNew->aLTerm[i] && pNew->aLTerm[i]->pExpr==pExpr ) nIn = 0; 2580 } 2581 }else if( ALWAYS(pExpr->x.pList && pExpr->x.pList->nExpr) ){ 2582 /* "x IN (value, value, ...)" */ 2583 nIn = sqlite3LogEst(pExpr->x.pList->nExpr); 2584 } 2585 if( pProbe->hasStat1 && rLogSize>=10 ){ 2586 LogEst M, logK, safetyMargin; 2587 /* Let: 2588 ** N = the total number of rows in the table 2589 ** K = the number of entries on the RHS of the IN operator 2590 ** M = the number of rows in the table that match terms to the 2591 ** to the left in the same index. If the IN operator is on 2592 ** the left-most index column, M==N. 2593 ** 2594 ** Given the definitions above, it is better to omit the IN operator 2595 ** from the index lookup and instead do a scan of the M elements, 2596 ** testing each scanned row against the IN operator separately, if: 2597 ** 2598 ** M*log(K) < K*log(N) 2599 ** 2600 ** Our estimates for M, K, and N might be inaccurate, so we build in 2601 ** a safety margin of 2 (LogEst: 10) that favors using the IN operator 2602 ** with the index, as using an index has better worst-case behavior. 2603 ** If we do not have real sqlite_stat1 data, always prefer to use 2604 ** the index. Do not bother with this optimization on very small 2605 ** tables (less than 2 rows) as it is pointless in that case. 2606 */ 2607 M = pProbe->aiRowLogEst[saved_nEq]; 2608 logK = estLog(nIn); 2609 safetyMargin = 10; /* TUNING: extra weight for indexed IN */ 2610 if( M + logK + safetyMargin < nIn + rLogSize ){ 2611 WHERETRACE(0x40, 2612 ("Scan preferred over IN operator on column %d of \"%s\" (%d<%d)\n", 2613 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2614 pNew->wsFlags |= WHERE_IN_SEEKSCAN; 2615 }else{ 2616 WHERETRACE(0x40, 2617 ("IN operator preferred on column %d of \"%s\" (%d>=%d)\n", 2618 saved_nEq, pProbe->zName, M+logK+10, nIn+rLogSize)); 2619 } 2620 } 2621 pNew->wsFlags |= WHERE_COLUMN_IN; 2622 }else if( eOp & (WO_EQ|WO_IS) ){ 2623 int iCol = pProbe->aiColumn[saved_nEq]; 2624 pNew->wsFlags |= WHERE_COLUMN_EQ; 2625 assert( saved_nEq==pNew->u.btree.nEq ); 2626 if( iCol==XN_ROWID 2627 || (iCol>=0 && nInMul==0 && saved_nEq==pProbe->nKeyCol-1) 2628 ){ 2629 if( iCol==XN_ROWID || pProbe->uniqNotNull 2630 || (pProbe->nKeyCol==1 && pProbe->onError && eOp==WO_EQ) 2631 ){ 2632 pNew->wsFlags |= WHERE_ONEROW; 2633 }else{ 2634 pNew->wsFlags |= WHERE_UNQ_WANTED; 2635 } 2636 } 2637 }else if( eOp & WO_ISNULL ){ 2638 pNew->wsFlags |= WHERE_COLUMN_NULL; 2639 }else if( eOp & (WO_GT|WO_GE) ){ 2640 testcase( eOp & WO_GT ); 2641 testcase( eOp & WO_GE ); 2642 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_BTM_LIMIT; 2643 pNew->u.btree.nBtm = whereRangeVectorLen( 2644 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2645 ); 2646 pBtm = pTerm; 2647 pTop = 0; 2648 if( pTerm->wtFlags & TERM_LIKEOPT ){ 2649 /* Range constraints that come from the LIKE optimization are 2650 ** always used in pairs. */ 2651 pTop = &pTerm[1]; 2652 assert( (pTop-(pTerm->pWC->a))<pTerm->pWC->nTerm ); 2653 assert( pTop->wtFlags & TERM_LIKEOPT ); 2654 assert( pTop->eOperator==WO_LT ); 2655 if( whereLoopResize(db, pNew, pNew->nLTerm+1) ) break; /* OOM */ 2656 pNew->aLTerm[pNew->nLTerm++] = pTop; 2657 pNew->wsFlags |= WHERE_TOP_LIMIT; 2658 pNew->u.btree.nTop = 1; 2659 } 2660 }else{ 2661 assert( eOp & (WO_LT|WO_LE) ); 2662 testcase( eOp & WO_LT ); 2663 testcase( eOp & WO_LE ); 2664 pNew->wsFlags |= WHERE_COLUMN_RANGE|WHERE_TOP_LIMIT; 2665 pNew->u.btree.nTop = whereRangeVectorLen( 2666 pParse, pSrc->iCursor, pProbe, saved_nEq, pTerm 2667 ); 2668 pTop = pTerm; 2669 pBtm = (pNew->wsFlags & WHERE_BTM_LIMIT)!=0 ? 2670 pNew->aLTerm[pNew->nLTerm-2] : 0; 2671 } 2672 2673 /* At this point pNew->nOut is set to the number of rows expected to 2674 ** be visited by the index scan before considering term pTerm, or the 2675 ** values of nIn and nInMul. In other words, assuming that all 2676 ** "x IN(...)" terms are replaced with "x = ?". This block updates 2677 ** the value of pNew->nOut to account for pTerm (but not nIn/nInMul). */ 2678 assert( pNew->nOut==saved_nOut ); 2679 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2680 /* Adjust nOut using stat4 data. Or, if there is no stat4 2681 ** data, using some other estimate. */ 2682 whereRangeScanEst(pParse, pBuilder, pBtm, pTop, pNew); 2683 }else{ 2684 int nEq = ++pNew->u.btree.nEq; 2685 assert( eOp & (WO_ISNULL|WO_EQ|WO_IN|WO_IS) ); 2686 2687 assert( pNew->nOut==saved_nOut ); 2688 if( pTerm->truthProb<=0 && pProbe->aiColumn[saved_nEq]>=0 ){ 2689 assert( (eOp & WO_IN) || nIn==0 ); 2690 testcase( eOp & WO_IN ); 2691 pNew->nOut += pTerm->truthProb; 2692 pNew->nOut -= nIn; 2693 }else{ 2694 #ifdef SQLITE_ENABLE_STAT4 2695 tRowcnt nOut = 0; 2696 if( nInMul==0 2697 && pProbe->nSample 2698 && pNew->u.btree.nEq<=pProbe->nSampleCol 2699 && ((eOp & WO_IN)==0 || !ExprHasProperty(pTerm->pExpr, EP_xIsSelect)) 2700 && OptimizationEnabled(db, SQLITE_Stat4) 2701 ){ 2702 Expr *pExpr = pTerm->pExpr; 2703 if( (eOp & (WO_EQ|WO_ISNULL|WO_IS))!=0 ){ 2704 testcase( eOp & WO_EQ ); 2705 testcase( eOp & WO_IS ); 2706 testcase( eOp & WO_ISNULL ); 2707 rc = whereEqualScanEst(pParse, pBuilder, pExpr->pRight, &nOut); 2708 }else{ 2709 rc = whereInScanEst(pParse, pBuilder, pExpr->x.pList, &nOut); 2710 } 2711 if( rc==SQLITE_NOTFOUND ) rc = SQLITE_OK; 2712 if( rc!=SQLITE_OK ) break; /* Jump out of the pTerm loop */ 2713 if( nOut ){ 2714 pNew->nOut = sqlite3LogEst(nOut); 2715 if( nEq==1 2716 /* TUNING: Mark terms as "low selectivity" if they seem likely 2717 ** to be true for half or more of the rows in the table. 2718 ** See tag-202002240-1 */ 2719 && pNew->nOut+10 > pProbe->aiRowLogEst[0] 2720 ){ 2721 #if WHERETRACE_ENABLED /* 0x01 */ 2722 if( sqlite3WhereTrace & 0x01 ){ 2723 sqlite3DebugPrintf( 2724 "STAT4 determines term has low selectivity:\n"); 2725 sqlite3WhereTermPrint(pTerm, 999); 2726 } 2727 #endif 2728 pTerm->wtFlags |= TERM_HIGHTRUTH; 2729 if( pTerm->wtFlags & TERM_HEURTRUTH ){ 2730 /* If the term has previously been used with an assumption of 2731 ** higher selectivity, then set the flag to rerun the 2732 ** loop computations. */ 2733 pBuilder->bldFlags2 |= SQLITE_BLDF2_2NDPASS; 2734 } 2735 } 2736 if( pNew->nOut>saved_nOut ) pNew->nOut = saved_nOut; 2737 pNew->nOut -= nIn; 2738 } 2739 } 2740 if( nOut==0 ) 2741 #endif 2742 { 2743 pNew->nOut += (pProbe->aiRowLogEst[nEq] - pProbe->aiRowLogEst[nEq-1]); 2744 if( eOp & WO_ISNULL ){ 2745 /* TUNING: If there is no likelihood() value, assume that a 2746 ** "col IS NULL" expression matches twice as many rows 2747 ** as (col=?). */ 2748 pNew->nOut += 10; 2749 } 2750 } 2751 } 2752 } 2753 2754 /* Set rCostIdx to the cost of visiting selected rows in index. Add 2755 ** it to pNew->rRun, which is currently set to the cost of the index 2756 ** seek only. Then, if this is a non-covering index, add the cost of 2757 ** visiting the rows in the main table. */ 2758 assert( pSrc->pTab->szTabRow>0 ); 2759 rCostIdx = pNew->nOut + 1 + (15*pProbe->szIdxRow)/pSrc->pTab->szTabRow; 2760 pNew->rRun = sqlite3LogEstAdd(rLogSize, rCostIdx); 2761 if( (pNew->wsFlags & (WHERE_IDX_ONLY|WHERE_IPK))==0 ){ 2762 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, pNew->nOut + 16); 2763 } 2764 ApplyCostMultiplier(pNew->rRun, pProbe->pTable->costMult); 2765 2766 nOutUnadjusted = pNew->nOut; 2767 pNew->rRun += nInMul + nIn; 2768 pNew->nOut += nInMul + nIn; 2769 whereLoopOutputAdjust(pBuilder->pWC, pNew, rSize); 2770 rc = whereLoopInsert(pBuilder, pNew); 2771 2772 if( pNew->wsFlags & WHERE_COLUMN_RANGE ){ 2773 pNew->nOut = saved_nOut; 2774 }else{ 2775 pNew->nOut = nOutUnadjusted; 2776 } 2777 2778 if( (pNew->wsFlags & WHERE_TOP_LIMIT)==0 2779 && pNew->u.btree.nEq<pProbe->nColumn 2780 ){ 2781 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nInMul+nIn); 2782 } 2783 pNew->nOut = saved_nOut; 2784 #ifdef SQLITE_ENABLE_STAT4 2785 pBuilder->nRecValid = nRecValid; 2786 #endif 2787 } 2788 pNew->prereq = saved_prereq; 2789 pNew->u.btree.nEq = saved_nEq; 2790 pNew->u.btree.nBtm = saved_nBtm; 2791 pNew->u.btree.nTop = saved_nTop; 2792 pNew->nSkip = saved_nSkip; 2793 pNew->wsFlags = saved_wsFlags; 2794 pNew->nOut = saved_nOut; 2795 pNew->nLTerm = saved_nLTerm; 2796 2797 /* Consider using a skip-scan if there are no WHERE clause constraints 2798 ** available for the left-most terms of the index, and if the average 2799 ** number of repeats in the left-most terms is at least 18. 2800 ** 2801 ** The magic number 18 is selected on the basis that scanning 17 rows 2802 ** is almost always quicker than an index seek (even though if the index 2803 ** contains fewer than 2^17 rows we assume otherwise in other parts of 2804 ** the code). And, even if it is not, it should not be too much slower. 2805 ** On the other hand, the extra seeks could end up being significantly 2806 ** more expensive. */ 2807 assert( 42==sqlite3LogEst(18) ); 2808 if( saved_nEq==saved_nSkip 2809 && saved_nEq+1<pProbe->nKeyCol 2810 && saved_nEq==pNew->nLTerm 2811 && pProbe->noSkipScan==0 2812 && pProbe->hasStat1!=0 2813 && OptimizationEnabled(db, SQLITE_SkipScan) 2814 && pProbe->aiRowLogEst[saved_nEq+1]>=42 /* TUNING: Minimum for skip-scan */ 2815 && (rc = whereLoopResize(db, pNew, pNew->nLTerm+1))==SQLITE_OK 2816 ){ 2817 LogEst nIter; 2818 pNew->u.btree.nEq++; 2819 pNew->nSkip++; 2820 pNew->aLTerm[pNew->nLTerm++] = 0; 2821 pNew->wsFlags |= WHERE_SKIPSCAN; 2822 nIter = pProbe->aiRowLogEst[saved_nEq] - pProbe->aiRowLogEst[saved_nEq+1]; 2823 pNew->nOut -= nIter; 2824 /* TUNING: Because uncertainties in the estimates for skip-scan queries, 2825 ** add a 1.375 fudge factor to make skip-scan slightly less likely. */ 2826 nIter += 5; 2827 whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, nIter + nInMul); 2828 pNew->nOut = saved_nOut; 2829 pNew->u.btree.nEq = saved_nEq; 2830 pNew->nSkip = saved_nSkip; 2831 pNew->wsFlags = saved_wsFlags; 2832 } 2833 2834 WHERETRACE(0x800, ("END %s.addBtreeIdx(%s), nEq=%d, rc=%d\n", 2835 pProbe->pTable->zName, pProbe->zName, saved_nEq, rc)); 2836 return rc; 2837 } 2838 2839 /* 2840 ** Return True if it is possible that pIndex might be useful in 2841 ** implementing the ORDER BY clause in pBuilder. 2842 ** 2843 ** Return False if pBuilder does not contain an ORDER BY clause or 2844 ** if there is no way for pIndex to be useful in implementing that 2845 ** ORDER BY clause. 2846 */ 2847 static int indexMightHelpWithOrderBy( 2848 WhereLoopBuilder *pBuilder, 2849 Index *pIndex, 2850 int iCursor 2851 ){ 2852 ExprList *pOB; 2853 ExprList *aColExpr; 2854 int ii, jj; 2855 2856 if( pIndex->bUnordered ) return 0; 2857 if( (pOB = pBuilder->pWInfo->pOrderBy)==0 ) return 0; 2858 for(ii=0; ii<pOB->nExpr; ii++){ 2859 Expr *pExpr = sqlite3ExprSkipCollateAndLikely(pOB->a[ii].pExpr); 2860 if( NEVER(pExpr==0) ) continue; 2861 if( pExpr->op==TK_COLUMN && pExpr->iTable==iCursor ){ 2862 if( pExpr->iColumn<0 ) return 1; 2863 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2864 if( pExpr->iColumn==pIndex->aiColumn[jj] ) return 1; 2865 } 2866 }else if( (aColExpr = pIndex->aColExpr)!=0 ){ 2867 for(jj=0; jj<pIndex->nKeyCol; jj++){ 2868 if( pIndex->aiColumn[jj]!=XN_EXPR ) continue; 2869 if( sqlite3ExprCompareSkip(pExpr,aColExpr->a[jj].pExpr,iCursor)==0 ){ 2870 return 1; 2871 } 2872 } 2873 } 2874 } 2875 return 0; 2876 } 2877 2878 /* Check to see if a partial index with pPartIndexWhere can be used 2879 ** in the current query. Return true if it can be and false if not. 2880 */ 2881 static int whereUsablePartialIndex( 2882 int iTab, /* The table for which we want an index */ 2883 int isLeft, /* True if iTab is the right table of a LEFT JOIN */ 2884 WhereClause *pWC, /* The WHERE clause of the query */ 2885 Expr *pWhere /* The WHERE clause from the partial index */ 2886 ){ 2887 int i; 2888 WhereTerm *pTerm; 2889 Parse *pParse = pWC->pWInfo->pParse; 2890 while( pWhere->op==TK_AND ){ 2891 if( !whereUsablePartialIndex(iTab,isLeft,pWC,pWhere->pLeft) ) return 0; 2892 pWhere = pWhere->pRight; 2893 } 2894 if( pParse->db->flags & SQLITE_EnableQPSG ) pParse = 0; 2895 for(i=0, pTerm=pWC->a; i<pWC->nTerm; i++, pTerm++){ 2896 Expr *pExpr; 2897 pExpr = pTerm->pExpr; 2898 if( (!ExprHasProperty(pExpr, EP_FromJoin) || pExpr->iRightJoinTable==iTab) 2899 && (isLeft==0 || ExprHasProperty(pExpr, EP_FromJoin)) 2900 && sqlite3ExprImpliesExpr(pParse, pExpr, pWhere, iTab) 2901 ){ 2902 return 1; 2903 } 2904 } 2905 return 0; 2906 } 2907 2908 /* 2909 ** Add all WhereLoop objects for a single table of the join where the table 2910 ** is identified by pBuilder->pNew->iTab. That table is guaranteed to be 2911 ** a b-tree table, not a virtual table. 2912 ** 2913 ** The costs (WhereLoop.rRun) of the b-tree loops added by this function 2914 ** are calculated as follows: 2915 ** 2916 ** For a full scan, assuming the table (or index) contains nRow rows: 2917 ** 2918 ** cost = nRow * 3.0 // full-table scan 2919 ** cost = nRow * K // scan of covering index 2920 ** cost = nRow * (K+3.0) // scan of non-covering index 2921 ** 2922 ** where K is a value between 1.1 and 3.0 set based on the relative 2923 ** estimated average size of the index and table records. 2924 ** 2925 ** For an index scan, where nVisit is the number of index rows visited 2926 ** by the scan, and nSeek is the number of seek operations required on 2927 ** the index b-tree: 2928 ** 2929 ** cost = nSeek * (log(nRow) + K * nVisit) // covering index 2930 ** cost = nSeek * (log(nRow) + (K+3.0) * nVisit) // non-covering index 2931 ** 2932 ** Normally, nSeek is 1. nSeek values greater than 1 come about if the 2933 ** WHERE clause includes "x IN (....)" terms used in place of "x=?". Or when 2934 ** implicit "x IN (SELECT x FROM tbl)" terms are added for skip-scans. 2935 ** 2936 ** The estimated values (nRow, nVisit, nSeek) often contain a large amount 2937 ** of uncertainty. For this reason, scoring is designed to pick plans that 2938 ** "do the least harm" if the estimates are inaccurate. For example, a 2939 ** log(nRow) factor is omitted from a non-covering index scan in order to 2940 ** bias the scoring in favor of using an index, since the worst-case 2941 ** performance of using an index is far better than the worst-case performance 2942 ** of a full table scan. 2943 */ 2944 static int whereLoopAddBtree( 2945 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 2946 Bitmask mPrereq /* Extra prerequesites for using this table */ 2947 ){ 2948 WhereInfo *pWInfo; /* WHERE analysis context */ 2949 Index *pProbe; /* An index we are evaluating */ 2950 Index sPk; /* A fake index object for the primary key */ 2951 LogEst aiRowEstPk[2]; /* The aiRowLogEst[] value for the sPk index */ 2952 i16 aiColumnPk = -1; /* The aColumn[] value for the sPk index */ 2953 SrcList *pTabList; /* The FROM clause */ 2954 SrcItem *pSrc; /* The FROM clause btree term to add */ 2955 WhereLoop *pNew; /* Template WhereLoop object */ 2956 int rc = SQLITE_OK; /* Return code */ 2957 int iSortIdx = 1; /* Index number */ 2958 int b; /* A boolean value */ 2959 LogEst rSize; /* number of rows in the table */ 2960 LogEst rLogSize; /* Logarithm of the number of rows in the table */ 2961 WhereClause *pWC; /* The parsed WHERE clause */ 2962 Table *pTab; /* Table being queried */ 2963 2964 pNew = pBuilder->pNew; 2965 pWInfo = pBuilder->pWInfo; 2966 pTabList = pWInfo->pTabList; 2967 pSrc = pTabList->a + pNew->iTab; 2968 pTab = pSrc->pTab; 2969 pWC = pBuilder->pWC; 2970 assert( !IsVirtual(pSrc->pTab) ); 2971 2972 if( pSrc->fg.isIndexedBy ){ 2973 /* An INDEXED BY clause specifies a particular index to use */ 2974 pProbe = pSrc->u2.pIBIndex; 2975 }else if( !HasRowid(pTab) ){ 2976 pProbe = pTab->pIndex; 2977 }else{ 2978 /* There is no INDEXED BY clause. Create a fake Index object in local 2979 ** variable sPk to represent the rowid primary key index. Make this 2980 ** fake index the first in a chain of Index objects with all of the real 2981 ** indices to follow */ 2982 Index *pFirst; /* First of real indices on the table */ 2983 memset(&sPk, 0, sizeof(Index)); 2984 sPk.nKeyCol = 1; 2985 sPk.nColumn = 1; 2986 sPk.aiColumn = &aiColumnPk; 2987 sPk.aiRowLogEst = aiRowEstPk; 2988 sPk.onError = OE_Replace; 2989 sPk.pTable = pTab; 2990 sPk.szIdxRow = pTab->szTabRow; 2991 sPk.idxType = SQLITE_IDXTYPE_IPK; 2992 aiRowEstPk[0] = pTab->nRowLogEst; 2993 aiRowEstPk[1] = 0; 2994 pFirst = pSrc->pTab->pIndex; 2995 if( pSrc->fg.notIndexed==0 ){ 2996 /* The real indices of the table are only considered if the 2997 ** NOT INDEXED qualifier is omitted from the FROM clause */ 2998 sPk.pNext = pFirst; 2999 } 3000 pProbe = &sPk; 3001 } 3002 rSize = pTab->nRowLogEst; 3003 rLogSize = estLog(rSize); 3004 3005 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 3006 /* Automatic indexes */ 3007 if( !pBuilder->pOrSet /* Not part of an OR optimization */ 3008 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 3009 && (pWInfo->pParse->db->flags & SQLITE_AutoIndex)!=0 3010 && !pSrc->fg.isIndexedBy /* Has no INDEXED BY clause */ 3011 && !pSrc->fg.notIndexed /* Has no NOT INDEXED clause */ 3012 && HasRowid(pTab) /* Not WITHOUT ROWID table. (FIXME: Why not?) */ 3013 && !pSrc->fg.isCorrelated /* Not a correlated subquery */ 3014 && !pSrc->fg.isRecursive /* Not a recursive common table expression. */ 3015 ){ 3016 /* Generate auto-index WhereLoops */ 3017 WhereTerm *pTerm; 3018 WhereTerm *pWCEnd = pWC->a + pWC->nTerm; 3019 for(pTerm=pWC->a; rc==SQLITE_OK && pTerm<pWCEnd; pTerm++){ 3020 if( pTerm->prereqRight & pNew->maskSelf ) continue; 3021 if( termCanDriveIndex(pTerm, pSrc, 0) ){ 3022 pNew->u.btree.nEq = 1; 3023 pNew->nSkip = 0; 3024 pNew->u.btree.pIndex = 0; 3025 pNew->nLTerm = 1; 3026 pNew->aLTerm[0] = pTerm; 3027 /* TUNING: One-time cost for computing the automatic index is 3028 ** estimated to be X*N*log2(N) where N is the number of rows in 3029 ** the table being indexed and where X is 7 (LogEst=28) for normal 3030 ** tables or 0.5 (LogEst=-10) for views and subqueries. The value 3031 ** of X is smaller for views and subqueries so that the query planner 3032 ** will be more aggressive about generating automatic indexes for 3033 ** those objects, since there is no opportunity to add schema 3034 ** indexes on subqueries and views. */ 3035 pNew->rSetup = rLogSize + rSize; 3036 if( pTab->pSelect==0 && (pTab->tabFlags & TF_Ephemeral)==0 ){ 3037 pNew->rSetup += 28; 3038 }else{ 3039 pNew->rSetup -= 10; 3040 } 3041 ApplyCostMultiplier(pNew->rSetup, pTab->costMult); 3042 if( pNew->rSetup<0 ) pNew->rSetup = 0; 3043 /* TUNING: Each index lookup yields 20 rows in the table. This 3044 ** is more than the usual guess of 10 rows, since we have no way 3045 ** of knowing how selective the index will ultimately be. It would 3046 ** not be unreasonable to make this value much larger. */ 3047 pNew->nOut = 43; assert( 43==sqlite3LogEst(20) ); 3048 pNew->rRun = sqlite3LogEstAdd(rLogSize,pNew->nOut); 3049 pNew->wsFlags = WHERE_AUTO_INDEX; 3050 pNew->prereq = mPrereq | pTerm->prereqRight; 3051 rc = whereLoopInsert(pBuilder, pNew); 3052 } 3053 } 3054 } 3055 #endif /* SQLITE_OMIT_AUTOMATIC_INDEX */ 3056 3057 /* Loop over all indices. If there was an INDEXED BY clause, then only 3058 ** consider index pProbe. */ 3059 for(; rc==SQLITE_OK && pProbe; 3060 pProbe=(pSrc->fg.isIndexedBy ? 0 : pProbe->pNext), iSortIdx++ 3061 ){ 3062 int isLeft = (pSrc->fg.jointype & JT_OUTER)!=0; 3063 if( pProbe->pPartIdxWhere!=0 3064 && !whereUsablePartialIndex(pSrc->iCursor, isLeft, pWC, 3065 pProbe->pPartIdxWhere) 3066 ){ 3067 testcase( pNew->iTab!=pSrc->iCursor ); /* See ticket [98d973b8f5] */ 3068 continue; /* Partial index inappropriate for this query */ 3069 } 3070 if( pProbe->bNoQuery ) continue; 3071 rSize = pProbe->aiRowLogEst[0]; 3072 pNew->u.btree.nEq = 0; 3073 pNew->u.btree.nBtm = 0; 3074 pNew->u.btree.nTop = 0; 3075 pNew->nSkip = 0; 3076 pNew->nLTerm = 0; 3077 pNew->iSortIdx = 0; 3078 pNew->rSetup = 0; 3079 pNew->prereq = mPrereq; 3080 pNew->nOut = rSize; 3081 pNew->u.btree.pIndex = pProbe; 3082 b = indexMightHelpWithOrderBy(pBuilder, pProbe, pSrc->iCursor); 3083 3084 /* The ONEPASS_DESIRED flags never occurs together with ORDER BY */ 3085 assert( (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || b==0 ); 3086 if( pProbe->idxType==SQLITE_IDXTYPE_IPK ){ 3087 /* Integer primary key index */ 3088 pNew->wsFlags = WHERE_IPK; 3089 3090 /* Full table scan */ 3091 pNew->iSortIdx = b ? iSortIdx : 0; 3092 /* TUNING: Cost of full table scan is 3.0*N. The 3.0 factor is an 3093 ** extra cost designed to discourage the use of full table scans, 3094 ** since index lookups have better worst-case performance if our 3095 ** stat guesses are wrong. Reduce the 3.0 penalty slightly 3096 ** (to 2.75) if we have valid STAT4 information for the table. 3097 ** At 2.75, a full table scan is preferred over using an index on 3098 ** a column with just two distinct values where each value has about 3099 ** an equal number of appearances. Without STAT4 data, we still want 3100 ** to use an index in that case, since the constraint might be for 3101 ** the scarcer of the two values, and in that case an index lookup is 3102 ** better. 3103 */ 3104 #ifdef SQLITE_ENABLE_STAT4 3105 pNew->rRun = rSize + 16 - 2*((pTab->tabFlags & TF_HasStat4)!=0); 3106 #else 3107 pNew->rRun = rSize + 16; 3108 #endif 3109 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3110 whereLoopOutputAdjust(pWC, pNew, rSize); 3111 rc = whereLoopInsert(pBuilder, pNew); 3112 pNew->nOut = rSize; 3113 if( rc ) break; 3114 }else{ 3115 Bitmask m; 3116 if( pProbe->isCovering ){ 3117 pNew->wsFlags = WHERE_IDX_ONLY | WHERE_INDEXED; 3118 m = 0; 3119 }else{ 3120 m = pSrc->colUsed & pProbe->colNotIdxed; 3121 pNew->wsFlags = (m==0) ? (WHERE_IDX_ONLY|WHERE_INDEXED) : WHERE_INDEXED; 3122 } 3123 3124 /* Full scan via index */ 3125 if( b 3126 || !HasRowid(pTab) 3127 || pProbe->pPartIdxWhere!=0 3128 || pSrc->fg.isIndexedBy 3129 || ( m==0 3130 && pProbe->bUnordered==0 3131 && (pProbe->szIdxRow<pTab->szTabRow) 3132 && (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 3133 && sqlite3GlobalConfig.bUseCis 3134 && OptimizationEnabled(pWInfo->pParse->db, SQLITE_CoverIdxScan) 3135 ) 3136 ){ 3137 pNew->iSortIdx = b ? iSortIdx : 0; 3138 3139 /* The cost of visiting the index rows is N*K, where K is 3140 ** between 1.1 and 3.0, depending on the relative sizes of the 3141 ** index and table rows. */ 3142 pNew->rRun = rSize + 1 + (15*pProbe->szIdxRow)/pTab->szTabRow; 3143 if( m!=0 ){ 3144 /* If this is a non-covering index scan, add in the cost of 3145 ** doing table lookups. The cost will be 3x the number of 3146 ** lookups. Take into account WHERE clause terms that can be 3147 ** satisfied using just the index, and that do not require a 3148 ** table lookup. */ 3149 LogEst nLookup = rSize + 16; /* Base cost: N*3 */ 3150 int ii; 3151 int iCur = pSrc->iCursor; 3152 WhereClause *pWC2 = &pWInfo->sWC; 3153 for(ii=0; ii<pWC2->nTerm; ii++){ 3154 WhereTerm *pTerm = &pWC2->a[ii]; 3155 if( !sqlite3ExprCoveredByIndex(pTerm->pExpr, iCur, pProbe) ){ 3156 break; 3157 } 3158 /* pTerm can be evaluated using just the index. So reduce 3159 ** the expected number of table lookups accordingly */ 3160 if( pTerm->truthProb<=0 ){ 3161 nLookup += pTerm->truthProb; 3162 }else{ 3163 nLookup--; 3164 if( pTerm->eOperator & (WO_EQ|WO_IS) ) nLookup -= 19; 3165 } 3166 } 3167 3168 pNew->rRun = sqlite3LogEstAdd(pNew->rRun, nLookup); 3169 } 3170 ApplyCostMultiplier(pNew->rRun, pTab->costMult); 3171 whereLoopOutputAdjust(pWC, pNew, rSize); 3172 rc = whereLoopInsert(pBuilder, pNew); 3173 pNew->nOut = rSize; 3174 if( rc ) break; 3175 } 3176 } 3177 3178 pBuilder->bldFlags1 = 0; 3179 rc = whereLoopAddBtreeIndex(pBuilder, pSrc, pProbe, 0); 3180 if( pBuilder->bldFlags1==SQLITE_BLDF1_INDEXED ){ 3181 /* If a non-unique index is used, or if a prefix of the key for 3182 ** unique index is used (making the index functionally non-unique) 3183 ** then the sqlite_stat1 data becomes important for scoring the 3184 ** plan */ 3185 pTab->tabFlags |= TF_StatsUsed; 3186 } 3187 #ifdef SQLITE_ENABLE_STAT4 3188 sqlite3Stat4ProbeFree(pBuilder->pRec); 3189 pBuilder->nRecValid = 0; 3190 pBuilder->pRec = 0; 3191 #endif 3192 } 3193 return rc; 3194 } 3195 3196 #ifndef SQLITE_OMIT_VIRTUALTABLE 3197 3198 /* 3199 ** Argument pIdxInfo is already populated with all constraints that may 3200 ** be used by the virtual table identified by pBuilder->pNew->iTab. This 3201 ** function marks a subset of those constraints usable, invokes the 3202 ** xBestIndex method and adds the returned plan to pBuilder. 3203 ** 3204 ** A constraint is marked usable if: 3205 ** 3206 ** * Argument mUsable indicates that its prerequisites are available, and 3207 ** 3208 ** * It is not one of the operators specified in the mExclude mask passed 3209 ** as the fourth argument (which in practice is either WO_IN or 0). 3210 ** 3211 ** Argument mPrereq is a mask of tables that must be scanned before the 3212 ** virtual table in question. These are added to the plans prerequisites 3213 ** before it is added to pBuilder. 3214 ** 3215 ** Output parameter *pbIn is set to true if the plan added to pBuilder 3216 ** uses one or more WO_IN terms, or false otherwise. 3217 */ 3218 static int whereLoopAddVirtualOne( 3219 WhereLoopBuilder *pBuilder, 3220 Bitmask mPrereq, /* Mask of tables that must be used. */ 3221 Bitmask mUsable, /* Mask of usable tables */ 3222 u16 mExclude, /* Exclude terms using these operators */ 3223 sqlite3_index_info *pIdxInfo, /* Populated object for xBestIndex */ 3224 u16 mNoOmit, /* Do not omit these constraints */ 3225 int *pbIn /* OUT: True if plan uses an IN(...) op */ 3226 ){ 3227 WhereClause *pWC = pBuilder->pWC; 3228 struct sqlite3_index_constraint *pIdxCons; 3229 struct sqlite3_index_constraint_usage *pUsage = pIdxInfo->aConstraintUsage; 3230 int i; 3231 int mxTerm; 3232 int rc = SQLITE_OK; 3233 WhereLoop *pNew = pBuilder->pNew; 3234 Parse *pParse = pBuilder->pWInfo->pParse; 3235 SrcItem *pSrc = &pBuilder->pWInfo->pTabList->a[pNew->iTab]; 3236 int nConstraint = pIdxInfo->nConstraint; 3237 3238 assert( (mUsable & mPrereq)==mPrereq ); 3239 *pbIn = 0; 3240 pNew->prereq = mPrereq; 3241 3242 /* Set the usable flag on the subset of constraints identified by 3243 ** arguments mUsable and mExclude. */ 3244 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3245 for(i=0; i<nConstraint; i++, pIdxCons++){ 3246 WhereTerm *pTerm = &pWC->a[pIdxCons->iTermOffset]; 3247 pIdxCons->usable = 0; 3248 if( (pTerm->prereqRight & mUsable)==pTerm->prereqRight 3249 && (pTerm->eOperator & mExclude)==0 3250 ){ 3251 pIdxCons->usable = 1; 3252 } 3253 } 3254 3255 /* Initialize the output fields of the sqlite3_index_info structure */ 3256 memset(pUsage, 0, sizeof(pUsage[0])*nConstraint); 3257 assert( pIdxInfo->needToFreeIdxStr==0 ); 3258 pIdxInfo->idxStr = 0; 3259 pIdxInfo->idxNum = 0; 3260 pIdxInfo->orderByConsumed = 0; 3261 pIdxInfo->estimatedCost = SQLITE_BIG_DBL / (double)2; 3262 pIdxInfo->estimatedRows = 25; 3263 pIdxInfo->idxFlags = 0; 3264 pIdxInfo->colUsed = (sqlite3_int64)pSrc->colUsed; 3265 3266 /* Invoke the virtual table xBestIndex() method */ 3267 rc = vtabBestIndex(pParse, pSrc->pTab, pIdxInfo); 3268 if( rc ){ 3269 if( rc==SQLITE_CONSTRAINT ){ 3270 /* If the xBestIndex method returns SQLITE_CONSTRAINT, that means 3271 ** that the particular combination of parameters provided is unusable. 3272 ** Make no entries in the loop table. 3273 */ 3274 WHERETRACE(0xffff, (" ^^^^--- non-viable plan rejected!\n")); 3275 return SQLITE_OK; 3276 } 3277 return rc; 3278 } 3279 3280 mxTerm = -1; 3281 assert( pNew->nLSlot>=nConstraint ); 3282 for(i=0; i<nConstraint; i++) pNew->aLTerm[i] = 0; 3283 pNew->u.vtab.omitMask = 0; 3284 pIdxCons = *(struct sqlite3_index_constraint**)&pIdxInfo->aConstraint; 3285 for(i=0; i<nConstraint; i++, pIdxCons++){ 3286 int iTerm; 3287 if( (iTerm = pUsage[i].argvIndex - 1)>=0 ){ 3288 WhereTerm *pTerm; 3289 int j = pIdxCons->iTermOffset; 3290 if( iTerm>=nConstraint 3291 || j<0 3292 || j>=pWC->nTerm 3293 || pNew->aLTerm[iTerm]!=0 3294 || pIdxCons->usable==0 3295 ){ 3296 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3297 testcase( pIdxInfo->needToFreeIdxStr ); 3298 return SQLITE_ERROR; 3299 } 3300 testcase( iTerm==nConstraint-1 ); 3301 testcase( j==0 ); 3302 testcase( j==pWC->nTerm-1 ); 3303 pTerm = &pWC->a[j]; 3304 pNew->prereq |= pTerm->prereqRight; 3305 assert( iTerm<pNew->nLSlot ); 3306 pNew->aLTerm[iTerm] = pTerm; 3307 if( iTerm>mxTerm ) mxTerm = iTerm; 3308 testcase( iTerm==15 ); 3309 testcase( iTerm==16 ); 3310 if( pUsage[i].omit ){ 3311 if( i<16 && ((1<<i)&mNoOmit)==0 ){ 3312 testcase( i!=iTerm ); 3313 pNew->u.vtab.omitMask |= 1<<iTerm; 3314 }else{ 3315 testcase( i!=iTerm ); 3316 } 3317 } 3318 if( (pTerm->eOperator & WO_IN)!=0 ){ 3319 /* A virtual table that is constrained by an IN clause may not 3320 ** consume the ORDER BY clause because (1) the order of IN terms 3321 ** is not necessarily related to the order of output terms and 3322 ** (2) Multiple outputs from a single IN value will not merge 3323 ** together. */ 3324 pIdxInfo->orderByConsumed = 0; 3325 pIdxInfo->idxFlags &= ~SQLITE_INDEX_SCAN_UNIQUE; 3326 *pbIn = 1; assert( (mExclude & WO_IN)==0 ); 3327 } 3328 } 3329 } 3330 3331 pNew->nLTerm = mxTerm+1; 3332 for(i=0; i<=mxTerm; i++){ 3333 if( pNew->aLTerm[i]==0 ){ 3334 /* The non-zero argvIdx values must be contiguous. Raise an 3335 ** error if they are not */ 3336 sqlite3ErrorMsg(pParse,"%s.xBestIndex malfunction",pSrc->pTab->zName); 3337 testcase( pIdxInfo->needToFreeIdxStr ); 3338 return SQLITE_ERROR; 3339 } 3340 } 3341 assert( pNew->nLTerm<=pNew->nLSlot ); 3342 pNew->u.vtab.idxNum = pIdxInfo->idxNum; 3343 pNew->u.vtab.needFree = pIdxInfo->needToFreeIdxStr; 3344 pIdxInfo->needToFreeIdxStr = 0; 3345 pNew->u.vtab.idxStr = pIdxInfo->idxStr; 3346 pNew->u.vtab.isOrdered = (i8)(pIdxInfo->orderByConsumed ? 3347 pIdxInfo->nOrderBy : 0); 3348 pNew->rSetup = 0; 3349 pNew->rRun = sqlite3LogEstFromDouble(pIdxInfo->estimatedCost); 3350 pNew->nOut = sqlite3LogEst(pIdxInfo->estimatedRows); 3351 3352 /* Set the WHERE_ONEROW flag if the xBestIndex() method indicated 3353 ** that the scan will visit at most one row. Clear it otherwise. */ 3354 if( pIdxInfo->idxFlags & SQLITE_INDEX_SCAN_UNIQUE ){ 3355 pNew->wsFlags |= WHERE_ONEROW; 3356 }else{ 3357 pNew->wsFlags &= ~WHERE_ONEROW; 3358 } 3359 rc = whereLoopInsert(pBuilder, pNew); 3360 if( pNew->u.vtab.needFree ){ 3361 sqlite3_free(pNew->u.vtab.idxStr); 3362 pNew->u.vtab.needFree = 0; 3363 } 3364 WHERETRACE(0xffff, (" bIn=%d prereqIn=%04llx prereqOut=%04llx\n", 3365 *pbIn, (sqlite3_uint64)mPrereq, 3366 (sqlite3_uint64)(pNew->prereq & ~mPrereq))); 3367 3368 return rc; 3369 } 3370 3371 /* 3372 ** If this function is invoked from within an xBestIndex() callback, it 3373 ** returns a pointer to a buffer containing the name of the collation 3374 ** sequence associated with element iCons of the sqlite3_index_info.aConstraint 3375 ** array. Or, if iCons is out of range or there is no active xBestIndex 3376 ** call, return NULL. 3377 */ 3378 const char *sqlite3_vtab_collation(sqlite3_index_info *pIdxInfo, int iCons){ 3379 HiddenIndexInfo *pHidden = (HiddenIndexInfo*)&pIdxInfo[1]; 3380 const char *zRet = 0; 3381 if( iCons>=0 && iCons<pIdxInfo->nConstraint ){ 3382 CollSeq *pC = 0; 3383 int iTerm = pIdxInfo->aConstraint[iCons].iTermOffset; 3384 Expr *pX = pHidden->pWC->a[iTerm].pExpr; 3385 if( pX->pLeft ){ 3386 pC = sqlite3ExprCompareCollSeq(pHidden->pParse, pX); 3387 } 3388 zRet = (pC ? pC->zName : sqlite3StrBINARY); 3389 } 3390 return zRet; 3391 } 3392 3393 /* 3394 ** Add all WhereLoop objects for a table of the join identified by 3395 ** pBuilder->pNew->iTab. That table is guaranteed to be a virtual table. 3396 ** 3397 ** If there are no LEFT or CROSS JOIN joins in the query, both mPrereq and 3398 ** mUnusable are set to 0. Otherwise, mPrereq is a mask of all FROM clause 3399 ** entries that occur before the virtual table in the FROM clause and are 3400 ** separated from it by at least one LEFT or CROSS JOIN. Similarly, the 3401 ** mUnusable mask contains all FROM clause entries that occur after the 3402 ** virtual table and are separated from it by at least one LEFT or 3403 ** CROSS JOIN. 3404 ** 3405 ** For example, if the query were: 3406 ** 3407 ** ... FROM t1, t2 LEFT JOIN t3, t4, vt CROSS JOIN t5, t6; 3408 ** 3409 ** then mPrereq corresponds to (t1, t2) and mUnusable to (t5, t6). 3410 ** 3411 ** All the tables in mPrereq must be scanned before the current virtual 3412 ** table. So any terms for which all prerequisites are satisfied by 3413 ** mPrereq may be specified as "usable" in all calls to xBestIndex. 3414 ** Conversely, all tables in mUnusable must be scanned after the current 3415 ** virtual table, so any terms for which the prerequisites overlap with 3416 ** mUnusable should always be configured as "not-usable" for xBestIndex. 3417 */ 3418 static int whereLoopAddVirtual( 3419 WhereLoopBuilder *pBuilder, /* WHERE clause information */ 3420 Bitmask mPrereq, /* Tables that must be scanned before this one */ 3421 Bitmask mUnusable /* Tables that must be scanned after this one */ 3422 ){ 3423 int rc = SQLITE_OK; /* Return code */ 3424 WhereInfo *pWInfo; /* WHERE analysis context */ 3425 Parse *pParse; /* The parsing context */ 3426 WhereClause *pWC; /* The WHERE clause */ 3427 SrcItem *pSrc; /* The FROM clause term to search */ 3428 sqlite3_index_info *p; /* Object to pass to xBestIndex() */ 3429 int nConstraint; /* Number of constraints in p */ 3430 int bIn; /* True if plan uses IN(...) operator */ 3431 WhereLoop *pNew; 3432 Bitmask mBest; /* Tables used by best possible plan */ 3433 u16 mNoOmit; 3434 3435 assert( (mPrereq & mUnusable)==0 ); 3436 pWInfo = pBuilder->pWInfo; 3437 pParse = pWInfo->pParse; 3438 pWC = pBuilder->pWC; 3439 pNew = pBuilder->pNew; 3440 pSrc = &pWInfo->pTabList->a[pNew->iTab]; 3441 assert( IsVirtual(pSrc->pTab) ); 3442 p = allocateIndexInfo(pParse, pWC, mUnusable, pSrc, pBuilder->pOrderBy, 3443 &mNoOmit); 3444 if( p==0 ) return SQLITE_NOMEM_BKPT; 3445 pNew->rSetup = 0; 3446 pNew->wsFlags = WHERE_VIRTUALTABLE; 3447 pNew->nLTerm = 0; 3448 pNew->u.vtab.needFree = 0; 3449 nConstraint = p->nConstraint; 3450 if( whereLoopResize(pParse->db, pNew, nConstraint) ){ 3451 sqlite3DbFree(pParse->db, p); 3452 return SQLITE_NOMEM_BKPT; 3453 } 3454 3455 /* First call xBestIndex() with all constraints usable. */ 3456 WHERETRACE(0x800, ("BEGIN %s.addVirtual()\n", pSrc->pTab->zName)); 3457 WHERETRACE(0x40, (" VirtualOne: all usable\n")); 3458 rc = whereLoopAddVirtualOne(pBuilder, mPrereq, ALLBITS, 0, p, mNoOmit, &bIn); 3459 3460 /* If the call to xBestIndex() with all terms enabled produced a plan 3461 ** that does not require any source tables (IOW: a plan with mBest==0) 3462 ** and does not use an IN(...) operator, then there is no point in making 3463 ** any further calls to xBestIndex() since they will all return the same 3464 ** result (if the xBestIndex() implementation is sane). */ 3465 if( rc==SQLITE_OK && ((mBest = (pNew->prereq & ~mPrereq))!=0 || bIn) ){ 3466 int seenZero = 0; /* True if a plan with no prereqs seen */ 3467 int seenZeroNoIN = 0; /* Plan with no prereqs and no IN(...) seen */ 3468 Bitmask mPrev = 0; 3469 Bitmask mBestNoIn = 0; 3470 3471 /* If the plan produced by the earlier call uses an IN(...) term, call 3472 ** xBestIndex again, this time with IN(...) terms disabled. */ 3473 if( bIn ){ 3474 WHERETRACE(0x40, (" VirtualOne: all usable w/o IN\n")); 3475 rc = whereLoopAddVirtualOne( 3476 pBuilder, mPrereq, ALLBITS, WO_IN, p, mNoOmit, &bIn); 3477 assert( bIn==0 ); 3478 mBestNoIn = pNew->prereq & ~mPrereq; 3479 if( mBestNoIn==0 ){ 3480 seenZero = 1; 3481 seenZeroNoIN = 1; 3482 } 3483 } 3484 3485 /* Call xBestIndex once for each distinct value of (prereqRight & ~mPrereq) 3486 ** in the set of terms that apply to the current virtual table. */ 3487 while( rc==SQLITE_OK ){ 3488 int i; 3489 Bitmask mNext = ALLBITS; 3490 assert( mNext>0 ); 3491 for(i=0; i<nConstraint; i++){ 3492 Bitmask mThis = ( 3493 pWC->a[p->aConstraint[i].iTermOffset].prereqRight & ~mPrereq 3494 ); 3495 if( mThis>mPrev && mThis<mNext ) mNext = mThis; 3496 } 3497 mPrev = mNext; 3498 if( mNext==ALLBITS ) break; 3499 if( mNext==mBest || mNext==mBestNoIn ) continue; 3500 WHERETRACE(0x40, (" VirtualOne: mPrev=%04llx mNext=%04llx\n", 3501 (sqlite3_uint64)mPrev, (sqlite3_uint64)mNext)); 3502 rc = whereLoopAddVirtualOne( 3503 pBuilder, mPrereq, mNext|mPrereq, 0, p, mNoOmit, &bIn); 3504 if( pNew->prereq==mPrereq ){ 3505 seenZero = 1; 3506 if( bIn==0 ) seenZeroNoIN = 1; 3507 } 3508 } 3509 3510 /* If the calls to xBestIndex() in the above loop did not find a plan 3511 ** that requires no source tables at all (i.e. one guaranteed to be 3512 ** usable), make a call here with all source tables disabled */ 3513 if( rc==SQLITE_OK && seenZero==0 ){ 3514 WHERETRACE(0x40, (" VirtualOne: all disabled\n")); 3515 rc = whereLoopAddVirtualOne( 3516 pBuilder, mPrereq, mPrereq, 0, p, mNoOmit, &bIn); 3517 if( bIn==0 ) seenZeroNoIN = 1; 3518 } 3519 3520 /* If the calls to xBestIndex() have so far failed to find a plan 3521 ** that requires no source tables at all and does not use an IN(...) 3522 ** operator, make a final call to obtain one here. */ 3523 if( rc==SQLITE_OK && seenZeroNoIN==0 ){ 3524 WHERETRACE(0x40, (" VirtualOne: all disabled and w/o IN\n")); 3525 rc = whereLoopAddVirtualOne( 3526 pBuilder, mPrereq, mPrereq, WO_IN, p, mNoOmit, &bIn); 3527 } 3528 } 3529 3530 if( p->needToFreeIdxStr ) sqlite3_free(p->idxStr); 3531 sqlite3DbFreeNN(pParse->db, p); 3532 WHERETRACE(0x800, ("END %s.addVirtual(), rc=%d\n", pSrc->pTab->zName, rc)); 3533 return rc; 3534 } 3535 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3536 3537 /* 3538 ** Add WhereLoop entries to handle OR terms. This works for either 3539 ** btrees or virtual tables. 3540 */ 3541 static int whereLoopAddOr( 3542 WhereLoopBuilder *pBuilder, 3543 Bitmask mPrereq, 3544 Bitmask mUnusable 3545 ){ 3546 WhereInfo *pWInfo = pBuilder->pWInfo; 3547 WhereClause *pWC; 3548 WhereLoop *pNew; 3549 WhereTerm *pTerm, *pWCEnd; 3550 int rc = SQLITE_OK; 3551 int iCur; 3552 WhereClause tempWC; 3553 WhereLoopBuilder sSubBuild; 3554 WhereOrSet sSum, sCur; 3555 SrcItem *pItem; 3556 3557 pWC = pBuilder->pWC; 3558 pWCEnd = pWC->a + pWC->nTerm; 3559 pNew = pBuilder->pNew; 3560 memset(&sSum, 0, sizeof(sSum)); 3561 pItem = pWInfo->pTabList->a + pNew->iTab; 3562 iCur = pItem->iCursor; 3563 3564 for(pTerm=pWC->a; pTerm<pWCEnd && rc==SQLITE_OK; pTerm++){ 3565 if( (pTerm->eOperator & WO_OR)!=0 3566 && (pTerm->u.pOrInfo->indexable & pNew->maskSelf)!=0 3567 ){ 3568 WhereClause * const pOrWC = &pTerm->u.pOrInfo->wc; 3569 WhereTerm * const pOrWCEnd = &pOrWC->a[pOrWC->nTerm]; 3570 WhereTerm *pOrTerm; 3571 int once = 1; 3572 int i, j; 3573 3574 sSubBuild = *pBuilder; 3575 sSubBuild.pOrderBy = 0; 3576 sSubBuild.pOrSet = &sCur; 3577 3578 WHERETRACE(0x200, ("Begin processing OR-clause %p\n", pTerm)); 3579 for(pOrTerm=pOrWC->a; pOrTerm<pOrWCEnd; pOrTerm++){ 3580 if( (pOrTerm->eOperator & WO_AND)!=0 ){ 3581 sSubBuild.pWC = &pOrTerm->u.pAndInfo->wc; 3582 }else if( pOrTerm->leftCursor==iCur ){ 3583 tempWC.pWInfo = pWC->pWInfo; 3584 tempWC.pOuter = pWC; 3585 tempWC.op = TK_AND; 3586 tempWC.nTerm = 1; 3587 tempWC.a = pOrTerm; 3588 sSubBuild.pWC = &tempWC; 3589 }else{ 3590 continue; 3591 } 3592 sCur.n = 0; 3593 #ifdef WHERETRACE_ENABLED 3594 WHERETRACE(0x200, ("OR-term %d of %p has %d subterms:\n", 3595 (int)(pOrTerm-pOrWC->a), pTerm, sSubBuild.pWC->nTerm)); 3596 if( sqlite3WhereTrace & 0x400 ){ 3597 sqlite3WhereClausePrint(sSubBuild.pWC); 3598 } 3599 #endif 3600 #ifndef SQLITE_OMIT_VIRTUALTABLE 3601 if( IsVirtual(pItem->pTab) ){ 3602 rc = whereLoopAddVirtual(&sSubBuild, mPrereq, mUnusable); 3603 }else 3604 #endif 3605 { 3606 rc = whereLoopAddBtree(&sSubBuild, mPrereq); 3607 } 3608 if( rc==SQLITE_OK ){ 3609 rc = whereLoopAddOr(&sSubBuild, mPrereq, mUnusable); 3610 } 3611 assert( rc==SQLITE_OK || rc==SQLITE_DONE || sCur.n==0 3612 || rc==SQLITE_NOMEM ); 3613 testcase( rc==SQLITE_NOMEM && sCur.n>0 ); 3614 testcase( rc==SQLITE_DONE ); 3615 if( sCur.n==0 ){ 3616 sSum.n = 0; 3617 break; 3618 }else if( once ){ 3619 whereOrMove(&sSum, &sCur); 3620 once = 0; 3621 }else{ 3622 WhereOrSet sPrev; 3623 whereOrMove(&sPrev, &sSum); 3624 sSum.n = 0; 3625 for(i=0; i<sPrev.n; i++){ 3626 for(j=0; j<sCur.n; j++){ 3627 whereOrInsert(&sSum, sPrev.a[i].prereq | sCur.a[j].prereq, 3628 sqlite3LogEstAdd(sPrev.a[i].rRun, sCur.a[j].rRun), 3629 sqlite3LogEstAdd(sPrev.a[i].nOut, sCur.a[j].nOut)); 3630 } 3631 } 3632 } 3633 } 3634 pNew->nLTerm = 1; 3635 pNew->aLTerm[0] = pTerm; 3636 pNew->wsFlags = WHERE_MULTI_OR; 3637 pNew->rSetup = 0; 3638 pNew->iSortIdx = 0; 3639 memset(&pNew->u, 0, sizeof(pNew->u)); 3640 for(i=0; rc==SQLITE_OK && i<sSum.n; i++){ 3641 /* TUNING: Currently sSum.a[i].rRun is set to the sum of the costs 3642 ** of all sub-scans required by the OR-scan. However, due to rounding 3643 ** errors, it may be that the cost of the OR-scan is equal to its 3644 ** most expensive sub-scan. Add the smallest possible penalty 3645 ** (equivalent to multiplying the cost by 1.07) to ensure that 3646 ** this does not happen. Otherwise, for WHERE clauses such as the 3647 ** following where there is an index on "y": 3648 ** 3649 ** WHERE likelihood(x=?, 0.99) OR y=? 3650 ** 3651 ** the planner may elect to "OR" together a full-table scan and an 3652 ** index lookup. And other similarly odd results. */ 3653 pNew->rRun = sSum.a[i].rRun + 1; 3654 pNew->nOut = sSum.a[i].nOut; 3655 pNew->prereq = sSum.a[i].prereq; 3656 rc = whereLoopInsert(pBuilder, pNew); 3657 } 3658 WHERETRACE(0x200, ("End processing OR-clause %p\n", pTerm)); 3659 } 3660 } 3661 return rc; 3662 } 3663 3664 /* 3665 ** Add all WhereLoop objects for all tables 3666 */ 3667 static int whereLoopAddAll(WhereLoopBuilder *pBuilder){ 3668 WhereInfo *pWInfo = pBuilder->pWInfo; 3669 Bitmask mPrereq = 0; 3670 Bitmask mPrior = 0; 3671 int iTab; 3672 SrcList *pTabList = pWInfo->pTabList; 3673 SrcItem *pItem; 3674 SrcItem *pEnd = &pTabList->a[pWInfo->nLevel]; 3675 sqlite3 *db = pWInfo->pParse->db; 3676 int rc = SQLITE_OK; 3677 WhereLoop *pNew; 3678 3679 /* Loop over the tables in the join, from left to right */ 3680 pNew = pBuilder->pNew; 3681 whereLoopInit(pNew); 3682 pBuilder->iPlanLimit = SQLITE_QUERY_PLANNER_LIMIT; 3683 for(iTab=0, pItem=pTabList->a; pItem<pEnd; iTab++, pItem++){ 3684 Bitmask mUnusable = 0; 3685 pNew->iTab = iTab; 3686 pBuilder->iPlanLimit += SQLITE_QUERY_PLANNER_LIMIT_INCR; 3687 pNew->maskSelf = sqlite3WhereGetMask(&pWInfo->sMaskSet, pItem->iCursor); 3688 if( (pItem->fg.jointype & (JT_LEFT|JT_CROSS))!=0 ){ 3689 /* This condition is true when pItem is the FROM clause term on the 3690 ** right-hand-side of a LEFT or CROSS JOIN. */ 3691 mPrereq = mPrior; 3692 }else{ 3693 mPrereq = 0; 3694 } 3695 #ifndef SQLITE_OMIT_VIRTUALTABLE 3696 if( IsVirtual(pItem->pTab) ){ 3697 SrcItem *p; 3698 for(p=&pItem[1]; p<pEnd; p++){ 3699 if( mUnusable || (p->fg.jointype & (JT_LEFT|JT_CROSS)) ){ 3700 mUnusable |= sqlite3WhereGetMask(&pWInfo->sMaskSet, p->iCursor); 3701 } 3702 } 3703 rc = whereLoopAddVirtual(pBuilder, mPrereq, mUnusable); 3704 }else 3705 #endif /* SQLITE_OMIT_VIRTUALTABLE */ 3706 { 3707 rc = whereLoopAddBtree(pBuilder, mPrereq); 3708 } 3709 if( rc==SQLITE_OK && pBuilder->pWC->hasOr ){ 3710 rc = whereLoopAddOr(pBuilder, mPrereq, mUnusable); 3711 } 3712 mPrior |= pNew->maskSelf; 3713 if( rc || db->mallocFailed ){ 3714 if( rc==SQLITE_DONE ){ 3715 /* We hit the query planner search limit set by iPlanLimit */ 3716 sqlite3_log(SQLITE_WARNING, "abbreviated query algorithm search"); 3717 rc = SQLITE_OK; 3718 }else{ 3719 break; 3720 } 3721 } 3722 } 3723 3724 whereLoopClear(db, pNew); 3725 return rc; 3726 } 3727 3728 /* 3729 ** Examine a WherePath (with the addition of the extra WhereLoop of the 6th 3730 ** parameters) to see if it outputs rows in the requested ORDER BY 3731 ** (or GROUP BY) without requiring a separate sort operation. Return N: 3732 ** 3733 ** N>0: N terms of the ORDER BY clause are satisfied 3734 ** N==0: No terms of the ORDER BY clause are satisfied 3735 ** N<0: Unknown yet how many terms of ORDER BY might be satisfied. 3736 ** 3737 ** Note that processing for WHERE_GROUPBY and WHERE_DISTINCTBY is not as 3738 ** strict. With GROUP BY and DISTINCT the only requirement is that 3739 ** equivalent rows appear immediately adjacent to one another. GROUP BY 3740 ** and DISTINCT do not require rows to appear in any particular order as long 3741 ** as equivalent rows are grouped together. Thus for GROUP BY and DISTINCT 3742 ** the pOrderBy terms can be matched in any order. With ORDER BY, the 3743 ** pOrderBy terms must be matched in strict left-to-right order. 3744 */ 3745 static i8 wherePathSatisfiesOrderBy( 3746 WhereInfo *pWInfo, /* The WHERE clause */ 3747 ExprList *pOrderBy, /* ORDER BY or GROUP BY or DISTINCT clause to check */ 3748 WherePath *pPath, /* The WherePath to check */ 3749 u16 wctrlFlags, /* WHERE_GROUPBY or _DISTINCTBY or _ORDERBY_LIMIT */ 3750 u16 nLoop, /* Number of entries in pPath->aLoop[] */ 3751 WhereLoop *pLast, /* Add this WhereLoop to the end of pPath->aLoop[] */ 3752 Bitmask *pRevMask /* OUT: Mask of WhereLoops to run in reverse order */ 3753 ){ 3754 u8 revSet; /* True if rev is known */ 3755 u8 rev; /* Composite sort order */ 3756 u8 revIdx; /* Index sort order */ 3757 u8 isOrderDistinct; /* All prior WhereLoops are order-distinct */ 3758 u8 distinctColumns; /* True if the loop has UNIQUE NOT NULL columns */ 3759 u8 isMatch; /* iColumn matches a term of the ORDER BY clause */ 3760 u16 eqOpMask; /* Allowed equality operators */ 3761 u16 nKeyCol; /* Number of key columns in pIndex */ 3762 u16 nColumn; /* Total number of ordered columns in the index */ 3763 u16 nOrderBy; /* Number terms in the ORDER BY clause */ 3764 int iLoop; /* Index of WhereLoop in pPath being processed */ 3765 int i, j; /* Loop counters */ 3766 int iCur; /* Cursor number for current WhereLoop */ 3767 int iColumn; /* A column number within table iCur */ 3768 WhereLoop *pLoop = 0; /* Current WhereLoop being processed. */ 3769 WhereTerm *pTerm; /* A single term of the WHERE clause */ 3770 Expr *pOBExpr; /* An expression from the ORDER BY clause */ 3771 CollSeq *pColl; /* COLLATE function from an ORDER BY clause term */ 3772 Index *pIndex; /* The index associated with pLoop */ 3773 sqlite3 *db = pWInfo->pParse->db; /* Database connection */ 3774 Bitmask obSat = 0; /* Mask of ORDER BY terms satisfied so far */ 3775 Bitmask obDone; /* Mask of all ORDER BY terms */ 3776 Bitmask orderDistinctMask; /* Mask of all well-ordered loops */ 3777 Bitmask ready; /* Mask of inner loops */ 3778 3779 /* 3780 ** We say the WhereLoop is "one-row" if it generates no more than one 3781 ** row of output. A WhereLoop is one-row if all of the following are true: 3782 ** (a) All index columns match with WHERE_COLUMN_EQ. 3783 ** (b) The index is unique 3784 ** Any WhereLoop with an WHERE_COLUMN_EQ constraint on the rowid is one-row. 3785 ** Every one-row WhereLoop will have the WHERE_ONEROW bit set in wsFlags. 3786 ** 3787 ** We say the WhereLoop is "order-distinct" if the set of columns from 3788 ** that WhereLoop that are in the ORDER BY clause are different for every 3789 ** row of the WhereLoop. Every one-row WhereLoop is automatically 3790 ** order-distinct. A WhereLoop that has no columns in the ORDER BY clause 3791 ** is not order-distinct. To be order-distinct is not quite the same as being 3792 ** UNIQUE since a UNIQUE column or index can have multiple rows that 3793 ** are NULL and NULL values are equivalent for the purpose of order-distinct. 3794 ** To be order-distinct, the columns must be UNIQUE and NOT NULL. 3795 ** 3796 ** The rowid for a table is always UNIQUE and NOT NULL so whenever the 3797 ** rowid appears in the ORDER BY clause, the corresponding WhereLoop is 3798 ** automatically order-distinct. 3799 */ 3800 3801 assert( pOrderBy!=0 ); 3802 if( nLoop && OptimizationDisabled(db, SQLITE_OrderByIdxJoin) ) return 0; 3803 3804 nOrderBy = pOrderBy->nExpr; 3805 testcase( nOrderBy==BMS-1 ); 3806 if( nOrderBy>BMS-1 ) return 0; /* Cannot optimize overly large ORDER BYs */ 3807 isOrderDistinct = 1; 3808 obDone = MASKBIT(nOrderBy)-1; 3809 orderDistinctMask = 0; 3810 ready = 0; 3811 eqOpMask = WO_EQ | WO_IS | WO_ISNULL; 3812 if( wctrlFlags & (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MAX|WHERE_ORDERBY_MIN) ){ 3813 eqOpMask |= WO_IN; 3814 } 3815 for(iLoop=0; isOrderDistinct && obSat<obDone && iLoop<=nLoop; iLoop++){ 3816 if( iLoop>0 ) ready |= pLoop->maskSelf; 3817 if( iLoop<nLoop ){ 3818 pLoop = pPath->aLoop[iLoop]; 3819 if( wctrlFlags & WHERE_ORDERBY_LIMIT ) continue; 3820 }else{ 3821 pLoop = pLast; 3822 } 3823 if( pLoop->wsFlags & WHERE_VIRTUALTABLE ){ 3824 if( pLoop->u.vtab.isOrdered && (wctrlFlags & WHERE_DISTINCTBY)==0 ){ 3825 obSat = obDone; 3826 } 3827 break; 3828 }else if( wctrlFlags & WHERE_DISTINCTBY ){ 3829 pLoop->u.btree.nDistinctCol = 0; 3830 } 3831 iCur = pWInfo->pTabList->a[pLoop->iTab].iCursor; 3832 3833 /* Mark off any ORDER BY term X that is a column in the table of 3834 ** the current loop for which there is term in the WHERE 3835 ** clause of the form X IS NULL or X=? that reference only outer 3836 ** loops. 3837 */ 3838 for(i=0; i<nOrderBy; i++){ 3839 if( MASKBIT(i) & obSat ) continue; 3840 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 3841 if( NEVER(pOBExpr==0) ) continue; 3842 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 3843 if( pOBExpr->iTable!=iCur ) continue; 3844 pTerm = sqlite3WhereFindTerm(&pWInfo->sWC, iCur, pOBExpr->iColumn, 3845 ~ready, eqOpMask, 0); 3846 if( pTerm==0 ) continue; 3847 if( pTerm->eOperator==WO_IN ){ 3848 /* IN terms are only valid for sorting in the ORDER BY LIMIT 3849 ** optimization, and then only if they are actually used 3850 ** by the query plan */ 3851 assert( wctrlFlags & 3852 (WHERE_ORDERBY_LIMIT|WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX) ); 3853 for(j=0; j<pLoop->nLTerm && pTerm!=pLoop->aLTerm[j]; j++){} 3854 if( j>=pLoop->nLTerm ) continue; 3855 } 3856 if( (pTerm->eOperator&(WO_EQ|WO_IS))!=0 && pOBExpr->iColumn>=0 ){ 3857 Parse *pParse = pWInfo->pParse; 3858 CollSeq *pColl1 = sqlite3ExprNNCollSeq(pParse, pOrderBy->a[i].pExpr); 3859 CollSeq *pColl2 = sqlite3ExprCompareCollSeq(pParse, pTerm->pExpr); 3860 assert( pColl1 ); 3861 if( pColl2==0 || sqlite3StrICmp(pColl1->zName, pColl2->zName) ){ 3862 continue; 3863 } 3864 testcase( pTerm->pExpr->op==TK_IS ); 3865 } 3866 obSat |= MASKBIT(i); 3867 } 3868 3869 if( (pLoop->wsFlags & WHERE_ONEROW)==0 ){ 3870 if( pLoop->wsFlags & WHERE_IPK ){ 3871 pIndex = 0; 3872 nKeyCol = 0; 3873 nColumn = 1; 3874 }else if( (pIndex = pLoop->u.btree.pIndex)==0 || pIndex->bUnordered ){ 3875 return 0; 3876 }else{ 3877 nKeyCol = pIndex->nKeyCol; 3878 nColumn = pIndex->nColumn; 3879 assert( nColumn==nKeyCol+1 || !HasRowid(pIndex->pTable) ); 3880 assert( pIndex->aiColumn[nColumn-1]==XN_ROWID 3881 || !HasRowid(pIndex->pTable)); 3882 /* All relevant terms of the index must also be non-NULL in order 3883 ** for isOrderDistinct to be true. So the isOrderDistint value 3884 ** computed here might be a false positive. Corrections will be 3885 ** made at tag-20210426-1 below */ 3886 isOrderDistinct = IsUniqueIndex(pIndex) 3887 && (pLoop->wsFlags & WHERE_SKIPSCAN)==0; 3888 } 3889 3890 /* Loop through all columns of the index and deal with the ones 3891 ** that are not constrained by == or IN. 3892 */ 3893 rev = revSet = 0; 3894 distinctColumns = 0; 3895 for(j=0; j<nColumn; j++){ 3896 u8 bOnce = 1; /* True to run the ORDER BY search loop */ 3897 3898 assert( j>=pLoop->u.btree.nEq 3899 || (pLoop->aLTerm[j]==0)==(j<pLoop->nSkip) 3900 ); 3901 if( j<pLoop->u.btree.nEq && j>=pLoop->nSkip ){ 3902 u16 eOp = pLoop->aLTerm[j]->eOperator; 3903 3904 /* Skip over == and IS and ISNULL terms. (Also skip IN terms when 3905 ** doing WHERE_ORDERBY_LIMIT processing). Except, IS and ISNULL 3906 ** terms imply that the index is not UNIQUE NOT NULL in which case 3907 ** the loop need to be marked as not order-distinct because it can 3908 ** have repeated NULL rows. 3909 ** 3910 ** If the current term is a column of an ((?,?) IN (SELECT...)) 3911 ** expression for which the SELECT returns more than one column, 3912 ** check that it is the only column used by this loop. Otherwise, 3913 ** if it is one of two or more, none of the columns can be 3914 ** considered to match an ORDER BY term. 3915 */ 3916 if( (eOp & eqOpMask)!=0 ){ 3917 if( eOp & (WO_ISNULL|WO_IS) ){ 3918 testcase( eOp & WO_ISNULL ); 3919 testcase( eOp & WO_IS ); 3920 testcase( isOrderDistinct ); 3921 isOrderDistinct = 0; 3922 } 3923 continue; 3924 }else if( ALWAYS(eOp & WO_IN) ){ 3925 /* ALWAYS() justification: eOp is an equality operator due to the 3926 ** j<pLoop->u.btree.nEq constraint above. Any equality other 3927 ** than WO_IN is captured by the previous "if". So this one 3928 ** always has to be WO_IN. */ 3929 Expr *pX = pLoop->aLTerm[j]->pExpr; 3930 for(i=j+1; i<pLoop->u.btree.nEq; i++){ 3931 if( pLoop->aLTerm[i]->pExpr==pX ){ 3932 assert( (pLoop->aLTerm[i]->eOperator & WO_IN) ); 3933 bOnce = 0; 3934 break; 3935 } 3936 } 3937 } 3938 } 3939 3940 /* Get the column number in the table (iColumn) and sort order 3941 ** (revIdx) for the j-th column of the index. 3942 */ 3943 if( pIndex ){ 3944 iColumn = pIndex->aiColumn[j]; 3945 revIdx = pIndex->aSortOrder[j] & KEYINFO_ORDER_DESC; 3946 if( iColumn==pIndex->pTable->iPKey ) iColumn = XN_ROWID; 3947 }else{ 3948 iColumn = XN_ROWID; 3949 revIdx = 0; 3950 } 3951 3952 /* An unconstrained column that might be NULL means that this 3953 ** WhereLoop is not well-ordered. tag-20210426-1 3954 */ 3955 if( isOrderDistinct ){ 3956 if( iColumn>=0 3957 && j>=pLoop->u.btree.nEq 3958 && pIndex->pTable->aCol[iColumn].notNull==0 3959 ){ 3960 isOrderDistinct = 0; 3961 } 3962 if( iColumn==XN_EXPR ){ 3963 isOrderDistinct = 0; 3964 } 3965 } 3966 3967 /* Find the ORDER BY term that corresponds to the j-th column 3968 ** of the index and mark that ORDER BY term off 3969 */ 3970 isMatch = 0; 3971 for(i=0; bOnce && i<nOrderBy; i++){ 3972 if( MASKBIT(i) & obSat ) continue; 3973 pOBExpr = sqlite3ExprSkipCollateAndLikely(pOrderBy->a[i].pExpr); 3974 testcase( wctrlFlags & WHERE_GROUPBY ); 3975 testcase( wctrlFlags & WHERE_DISTINCTBY ); 3976 if( NEVER(pOBExpr==0) ) continue; 3977 if( (wctrlFlags & (WHERE_GROUPBY|WHERE_DISTINCTBY))==0 ) bOnce = 0; 3978 if( iColumn>=XN_ROWID ){ 3979 if( pOBExpr->op!=TK_COLUMN && pOBExpr->op!=TK_AGG_COLUMN ) continue; 3980 if( pOBExpr->iTable!=iCur ) continue; 3981 if( pOBExpr->iColumn!=iColumn ) continue; 3982 }else{ 3983 Expr *pIdxExpr = pIndex->aColExpr->a[j].pExpr; 3984 if( sqlite3ExprCompareSkip(pOBExpr, pIdxExpr, iCur) ){ 3985 continue; 3986 } 3987 } 3988 if( iColumn!=XN_ROWID ){ 3989 pColl = sqlite3ExprNNCollSeq(pWInfo->pParse, pOrderBy->a[i].pExpr); 3990 if( sqlite3StrICmp(pColl->zName, pIndex->azColl[j])!=0 ) continue; 3991 } 3992 if( wctrlFlags & WHERE_DISTINCTBY ){ 3993 pLoop->u.btree.nDistinctCol = j+1; 3994 } 3995 isMatch = 1; 3996 break; 3997 } 3998 if( isMatch && (wctrlFlags & WHERE_GROUPBY)==0 ){ 3999 /* Make sure the sort order is compatible in an ORDER BY clause. 4000 ** Sort order is irrelevant for a GROUP BY clause. */ 4001 if( revSet ){ 4002 if( (rev ^ revIdx)!=(pOrderBy->a[i].sortFlags&KEYINFO_ORDER_DESC) ){ 4003 isMatch = 0; 4004 } 4005 }else{ 4006 rev = revIdx ^ (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_DESC); 4007 if( rev ) *pRevMask |= MASKBIT(iLoop); 4008 revSet = 1; 4009 } 4010 } 4011 if( isMatch && (pOrderBy->a[i].sortFlags & KEYINFO_ORDER_BIGNULL) ){ 4012 if( j==pLoop->u.btree.nEq ){ 4013 pLoop->wsFlags |= WHERE_BIGNULL_SORT; 4014 }else{ 4015 isMatch = 0; 4016 } 4017 } 4018 if( isMatch ){ 4019 if( iColumn==XN_ROWID ){ 4020 testcase( distinctColumns==0 ); 4021 distinctColumns = 1; 4022 } 4023 obSat |= MASKBIT(i); 4024 }else{ 4025 /* No match found */ 4026 if( j==0 || j<nKeyCol ){ 4027 testcase( isOrderDistinct!=0 ); 4028 isOrderDistinct = 0; 4029 } 4030 break; 4031 } 4032 } /* end Loop over all index columns */ 4033 if( distinctColumns ){ 4034 testcase( isOrderDistinct==0 ); 4035 isOrderDistinct = 1; 4036 } 4037 } /* end-if not one-row */ 4038 4039 /* Mark off any other ORDER BY terms that reference pLoop */ 4040 if( isOrderDistinct ){ 4041 orderDistinctMask |= pLoop->maskSelf; 4042 for(i=0; i<nOrderBy; i++){ 4043 Expr *p; 4044 Bitmask mTerm; 4045 if( MASKBIT(i) & obSat ) continue; 4046 p = pOrderBy->a[i].pExpr; 4047 mTerm = sqlite3WhereExprUsage(&pWInfo->sMaskSet,p); 4048 if( mTerm==0 && !sqlite3ExprIsConstant(p) ) continue; 4049 if( (mTerm&~orderDistinctMask)==0 ){ 4050 obSat |= MASKBIT(i); 4051 } 4052 } 4053 } 4054 } /* End the loop over all WhereLoops from outer-most down to inner-most */ 4055 if( obSat==obDone ) return (i8)nOrderBy; 4056 if( !isOrderDistinct ){ 4057 for(i=nOrderBy-1; i>0; i--){ 4058 Bitmask m = MASKBIT(i) - 1; 4059 if( (obSat&m)==m ) return i; 4060 } 4061 return 0; 4062 } 4063 return -1; 4064 } 4065 4066 4067 /* 4068 ** If the WHERE_GROUPBY flag is set in the mask passed to sqlite3WhereBegin(), 4069 ** the planner assumes that the specified pOrderBy list is actually a GROUP 4070 ** BY clause - and so any order that groups rows as required satisfies the 4071 ** request. 4072 ** 4073 ** Normally, in this case it is not possible for the caller to determine 4074 ** whether or not the rows are really being delivered in sorted order, or 4075 ** just in some other order that provides the required grouping. However, 4076 ** if the WHERE_SORTBYGROUP flag is also passed to sqlite3WhereBegin(), then 4077 ** this function may be called on the returned WhereInfo object. It returns 4078 ** true if the rows really will be sorted in the specified order, or false 4079 ** otherwise. 4080 ** 4081 ** For example, assuming: 4082 ** 4083 ** CREATE INDEX i1 ON t1(x, Y); 4084 ** 4085 ** then 4086 ** 4087 ** SELECT * FROM t1 GROUP BY x,y ORDER BY x,y; -- IsSorted()==1 4088 ** SELECT * FROM t1 GROUP BY y,x ORDER BY y,x; -- IsSorted()==0 4089 */ 4090 int sqlite3WhereIsSorted(WhereInfo *pWInfo){ 4091 assert( pWInfo->wctrlFlags & WHERE_GROUPBY ); 4092 assert( pWInfo->wctrlFlags & WHERE_SORTBYGROUP ); 4093 return pWInfo->sorted; 4094 } 4095 4096 #ifdef WHERETRACE_ENABLED 4097 /* For debugging use only: */ 4098 static const char *wherePathName(WherePath *pPath, int nLoop, WhereLoop *pLast){ 4099 static char zName[65]; 4100 int i; 4101 for(i=0; i<nLoop; i++){ zName[i] = pPath->aLoop[i]->cId; } 4102 if( pLast ) zName[i++] = pLast->cId; 4103 zName[i] = 0; 4104 return zName; 4105 } 4106 #endif 4107 4108 /* 4109 ** Return the cost of sorting nRow rows, assuming that the keys have 4110 ** nOrderby columns and that the first nSorted columns are already in 4111 ** order. 4112 */ 4113 static LogEst whereSortingCost( 4114 WhereInfo *pWInfo, 4115 LogEst nRow, 4116 int nOrderBy, 4117 int nSorted 4118 ){ 4119 /* TUNING: Estimated cost of a full external sort, where N is 4120 ** the number of rows to sort is: 4121 ** 4122 ** cost = (3.0 * N * log(N)). 4123 ** 4124 ** Or, if the order-by clause has X terms but only the last Y 4125 ** terms are out of order, then block-sorting will reduce the 4126 ** sorting cost to: 4127 ** 4128 ** cost = (3.0 * N * log(N)) * (Y/X) 4129 ** 4130 ** The (Y/X) term is implemented using stack variable rScale 4131 ** below. 4132 */ 4133 LogEst rScale, rSortCost; 4134 assert( nOrderBy>0 && 66==sqlite3LogEst(100) ); 4135 rScale = sqlite3LogEst((nOrderBy-nSorted)*100/nOrderBy) - 66; 4136 rSortCost = nRow + rScale + 16; 4137 4138 /* Multiple by log(M) where M is the number of output rows. 4139 ** Use the LIMIT for M if it is smaller. Or if this sort is for 4140 ** a DISTINCT operator, M will be the number of distinct output 4141 ** rows, so fudge it downwards a bit. 4142 */ 4143 if( (pWInfo->wctrlFlags & WHERE_USE_LIMIT)!=0 && pWInfo->iLimit<nRow ){ 4144 nRow = pWInfo->iLimit; 4145 }else if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT) ){ 4146 /* TUNING: In the sort for a DISTINCT operator, assume that the DISTINCT 4147 ** reduces the number of output rows by a factor of 2 */ 4148 if( nRow>10 ) nRow -= 10; assert( 10==sqlite3LogEst(2) ); 4149 } 4150 rSortCost += estLog(nRow); 4151 return rSortCost; 4152 } 4153 4154 /* 4155 ** Given the list of WhereLoop objects at pWInfo->pLoops, this routine 4156 ** attempts to find the lowest cost path that visits each WhereLoop 4157 ** once. This path is then loaded into the pWInfo->a[].pWLoop fields. 4158 ** 4159 ** Assume that the total number of output rows that will need to be sorted 4160 ** will be nRowEst (in the 10*log2 representation). Or, ignore sorting 4161 ** costs if nRowEst==0. 4162 ** 4163 ** Return SQLITE_OK on success or SQLITE_NOMEM of a memory allocation 4164 ** error occurs. 4165 */ 4166 static int wherePathSolver(WhereInfo *pWInfo, LogEst nRowEst){ 4167 int mxChoice; /* Maximum number of simultaneous paths tracked */ 4168 int nLoop; /* Number of terms in the join */ 4169 Parse *pParse; /* Parsing context */ 4170 sqlite3 *db; /* The database connection */ 4171 int iLoop; /* Loop counter over the terms of the join */ 4172 int ii, jj; /* Loop counters */ 4173 int mxI = 0; /* Index of next entry to replace */ 4174 int nOrderBy; /* Number of ORDER BY clause terms */ 4175 LogEst mxCost = 0; /* Maximum cost of a set of paths */ 4176 LogEst mxUnsorted = 0; /* Maximum unsorted cost of a set of path */ 4177 int nTo, nFrom; /* Number of valid entries in aTo[] and aFrom[] */ 4178 WherePath *aFrom; /* All nFrom paths at the previous level */ 4179 WherePath *aTo; /* The nTo best paths at the current level */ 4180 WherePath *pFrom; /* An element of aFrom[] that we are working on */ 4181 WherePath *pTo; /* An element of aTo[] that we are working on */ 4182 WhereLoop *pWLoop; /* One of the WhereLoop objects */ 4183 WhereLoop **pX; /* Used to divy up the pSpace memory */ 4184 LogEst *aSortCost = 0; /* Sorting and partial sorting costs */ 4185 char *pSpace; /* Temporary memory used by this routine */ 4186 int nSpace; /* Bytes of space allocated at pSpace */ 4187 4188 pParse = pWInfo->pParse; 4189 db = pParse->db; 4190 nLoop = pWInfo->nLevel; 4191 /* TUNING: For simple queries, only the best path is tracked. 4192 ** For 2-way joins, the 5 best paths are followed. 4193 ** For joins of 3 or more tables, track the 10 best paths */ 4194 mxChoice = (nLoop<=1) ? 1 : (nLoop==2 ? 5 : 10); 4195 assert( nLoop<=pWInfo->pTabList->nSrc ); 4196 WHERETRACE(0x002, ("---- begin solver. (nRowEst=%d)\n", nRowEst)); 4197 4198 /* If nRowEst is zero and there is an ORDER BY clause, ignore it. In this 4199 ** case the purpose of this call is to estimate the number of rows returned 4200 ** by the overall query. Once this estimate has been obtained, the caller 4201 ** will invoke this function a second time, passing the estimate as the 4202 ** nRowEst parameter. */ 4203 if( pWInfo->pOrderBy==0 || nRowEst==0 ){ 4204 nOrderBy = 0; 4205 }else{ 4206 nOrderBy = pWInfo->pOrderBy->nExpr; 4207 } 4208 4209 /* Allocate and initialize space for aTo, aFrom and aSortCost[] */ 4210 nSpace = (sizeof(WherePath)+sizeof(WhereLoop*)*nLoop)*mxChoice*2; 4211 nSpace += sizeof(LogEst) * nOrderBy; 4212 pSpace = sqlite3DbMallocRawNN(db, nSpace); 4213 if( pSpace==0 ) return SQLITE_NOMEM_BKPT; 4214 aTo = (WherePath*)pSpace; 4215 aFrom = aTo+mxChoice; 4216 memset(aFrom, 0, sizeof(aFrom[0])); 4217 pX = (WhereLoop**)(aFrom+mxChoice); 4218 for(ii=mxChoice*2, pFrom=aTo; ii>0; ii--, pFrom++, pX += nLoop){ 4219 pFrom->aLoop = pX; 4220 } 4221 if( nOrderBy ){ 4222 /* If there is an ORDER BY clause and it is not being ignored, set up 4223 ** space for the aSortCost[] array. Each element of the aSortCost array 4224 ** is either zero - meaning it has not yet been initialized - or the 4225 ** cost of sorting nRowEst rows of data where the first X terms of 4226 ** the ORDER BY clause are already in order, where X is the array 4227 ** index. */ 4228 aSortCost = (LogEst*)pX; 4229 memset(aSortCost, 0, sizeof(LogEst) * nOrderBy); 4230 } 4231 assert( aSortCost==0 || &pSpace[nSpace]==(char*)&aSortCost[nOrderBy] ); 4232 assert( aSortCost!=0 || &pSpace[nSpace]==(char*)pX ); 4233 4234 /* Seed the search with a single WherePath containing zero WhereLoops. 4235 ** 4236 ** TUNING: Do not let the number of iterations go above 28. If the cost 4237 ** of computing an automatic index is not paid back within the first 28 4238 ** rows, then do not use the automatic index. */ 4239 aFrom[0].nRow = MIN(pParse->nQueryLoop, 48); assert( 48==sqlite3LogEst(28) ); 4240 nFrom = 1; 4241 assert( aFrom[0].isOrdered==0 ); 4242 if( nOrderBy ){ 4243 /* If nLoop is zero, then there are no FROM terms in the query. Since 4244 ** in this case the query may return a maximum of one row, the results 4245 ** are already in the requested order. Set isOrdered to nOrderBy to 4246 ** indicate this. Or, if nLoop is greater than zero, set isOrdered to 4247 ** -1, indicating that the result set may or may not be ordered, 4248 ** depending on the loops added to the current plan. */ 4249 aFrom[0].isOrdered = nLoop>0 ? -1 : nOrderBy; 4250 } 4251 4252 /* Compute successively longer WherePaths using the previous generation 4253 ** of WherePaths as the basis for the next. Keep track of the mxChoice 4254 ** best paths at each generation */ 4255 for(iLoop=0; iLoop<nLoop; iLoop++){ 4256 nTo = 0; 4257 for(ii=0, pFrom=aFrom; ii<nFrom; ii++, pFrom++){ 4258 for(pWLoop=pWInfo->pLoops; pWLoop; pWLoop=pWLoop->pNextLoop){ 4259 LogEst nOut; /* Rows visited by (pFrom+pWLoop) */ 4260 LogEst rCost; /* Cost of path (pFrom+pWLoop) */ 4261 LogEst rUnsorted; /* Unsorted cost of (pFrom+pWLoop) */ 4262 i8 isOrdered = pFrom->isOrdered; /* isOrdered for (pFrom+pWLoop) */ 4263 Bitmask maskNew; /* Mask of src visited by (..) */ 4264 Bitmask revMask = 0; /* Mask of rev-order loops for (..) */ 4265 4266 if( (pWLoop->prereq & ~pFrom->maskLoop)!=0 ) continue; 4267 if( (pWLoop->maskSelf & pFrom->maskLoop)!=0 ) continue; 4268 if( (pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 && pFrom->nRow<3 ){ 4269 /* Do not use an automatic index if the this loop is expected 4270 ** to run less than 1.25 times. It is tempting to also exclude 4271 ** automatic index usage on an outer loop, but sometimes an automatic 4272 ** index is useful in the outer loop of a correlated subquery. */ 4273 assert( 10==sqlite3LogEst(2) ); 4274 continue; 4275 } 4276 4277 /* At this point, pWLoop is a candidate to be the next loop. 4278 ** Compute its cost */ 4279 rUnsorted = sqlite3LogEstAdd(pWLoop->rSetup,pWLoop->rRun + pFrom->nRow); 4280 rUnsorted = sqlite3LogEstAdd(rUnsorted, pFrom->rUnsorted); 4281 nOut = pFrom->nRow + pWLoop->nOut; 4282 maskNew = pFrom->maskLoop | pWLoop->maskSelf; 4283 if( isOrdered<0 ){ 4284 isOrdered = wherePathSatisfiesOrderBy(pWInfo, 4285 pWInfo->pOrderBy, pFrom, pWInfo->wctrlFlags, 4286 iLoop, pWLoop, &revMask); 4287 }else{ 4288 revMask = pFrom->revLoop; 4289 } 4290 if( isOrdered>=0 && isOrdered<nOrderBy ){ 4291 if( aSortCost[isOrdered]==0 ){ 4292 aSortCost[isOrdered] = whereSortingCost( 4293 pWInfo, nRowEst, nOrderBy, isOrdered 4294 ); 4295 } 4296 /* TUNING: Add a small extra penalty (5) to sorting as an 4297 ** extra encouragment to the query planner to select a plan 4298 ** where the rows emerge in the correct order without any sorting 4299 ** required. */ 4300 rCost = sqlite3LogEstAdd(rUnsorted, aSortCost[isOrdered]) + 5; 4301 4302 WHERETRACE(0x002, 4303 ("---- sort cost=%-3d (%d/%d) increases cost %3d to %-3d\n", 4304 aSortCost[isOrdered], (nOrderBy-isOrdered), nOrderBy, 4305 rUnsorted, rCost)); 4306 }else{ 4307 rCost = rUnsorted; 4308 rUnsorted -= 2; /* TUNING: Slight bias in favor of no-sort plans */ 4309 } 4310 4311 /* Check to see if pWLoop should be added to the set of 4312 ** mxChoice best-so-far paths. 4313 ** 4314 ** First look for an existing path among best-so-far paths 4315 ** that covers the same set of loops and has the same isOrdered 4316 ** setting as the current path candidate. 4317 ** 4318 ** The term "((pTo->isOrdered^isOrdered)&0x80)==0" is equivalent 4319 ** to (pTo->isOrdered==(-1))==(isOrdered==(-1))" for the range 4320 ** of legal values for isOrdered, -1..64. 4321 */ 4322 for(jj=0, pTo=aTo; jj<nTo; jj++, pTo++){ 4323 if( pTo->maskLoop==maskNew 4324 && ((pTo->isOrdered^isOrdered)&0x80)==0 4325 ){ 4326 testcase( jj==nTo-1 ); 4327 break; 4328 } 4329 } 4330 if( jj>=nTo ){ 4331 /* None of the existing best-so-far paths match the candidate. */ 4332 if( nTo>=mxChoice 4333 && (rCost>mxCost || (rCost==mxCost && rUnsorted>=mxUnsorted)) 4334 ){ 4335 /* The current candidate is no better than any of the mxChoice 4336 ** paths currently in the best-so-far buffer. So discard 4337 ** this candidate as not viable. */ 4338 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4339 if( sqlite3WhereTrace&0x4 ){ 4340 sqlite3DebugPrintf("Skip %s cost=%-3d,%3d,%3d order=%c\n", 4341 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4342 isOrdered>=0 ? isOrdered+'0' : '?'); 4343 } 4344 #endif 4345 continue; 4346 } 4347 /* If we reach this points it means that the new candidate path 4348 ** needs to be added to the set of best-so-far paths. */ 4349 if( nTo<mxChoice ){ 4350 /* Increase the size of the aTo set by one */ 4351 jj = nTo++; 4352 }else{ 4353 /* New path replaces the prior worst to keep count below mxChoice */ 4354 jj = mxI; 4355 } 4356 pTo = &aTo[jj]; 4357 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4358 if( sqlite3WhereTrace&0x4 ){ 4359 sqlite3DebugPrintf("New %s cost=%-3d,%3d,%3d order=%c\n", 4360 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4361 isOrdered>=0 ? isOrdered+'0' : '?'); 4362 } 4363 #endif 4364 }else{ 4365 /* Control reaches here if best-so-far path pTo=aTo[jj] covers the 4366 ** same set of loops and has the same isOrdered setting as the 4367 ** candidate path. Check to see if the candidate should replace 4368 ** pTo or if the candidate should be skipped. 4369 ** 4370 ** The conditional is an expanded vector comparison equivalent to: 4371 ** (pTo->rCost,pTo->nRow,pTo->rUnsorted) <= (rCost,nOut,rUnsorted) 4372 */ 4373 if( pTo->rCost<rCost 4374 || (pTo->rCost==rCost 4375 && (pTo->nRow<nOut 4376 || (pTo->nRow==nOut && pTo->rUnsorted<=rUnsorted) 4377 ) 4378 ) 4379 ){ 4380 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4381 if( sqlite3WhereTrace&0x4 ){ 4382 sqlite3DebugPrintf( 4383 "Skip %s cost=%-3d,%3d,%3d order=%c", 4384 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4385 isOrdered>=0 ? isOrdered+'0' : '?'); 4386 sqlite3DebugPrintf(" vs %s cost=%-3d,%3d,%3d order=%c\n", 4387 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4388 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4389 } 4390 #endif 4391 /* Discard the candidate path from further consideration */ 4392 testcase( pTo->rCost==rCost ); 4393 continue; 4394 } 4395 testcase( pTo->rCost==rCost+1 ); 4396 /* Control reaches here if the candidate path is better than the 4397 ** pTo path. Replace pTo with the candidate. */ 4398 #ifdef WHERETRACE_ENABLED /* 0x4 */ 4399 if( sqlite3WhereTrace&0x4 ){ 4400 sqlite3DebugPrintf( 4401 "Update %s cost=%-3d,%3d,%3d order=%c", 4402 wherePathName(pFrom, iLoop, pWLoop), rCost, nOut, rUnsorted, 4403 isOrdered>=0 ? isOrdered+'0' : '?'); 4404 sqlite3DebugPrintf(" was %s cost=%-3d,%3d,%3d order=%c\n", 4405 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4406 pTo->rUnsorted, pTo->isOrdered>=0 ? pTo->isOrdered+'0' : '?'); 4407 } 4408 #endif 4409 } 4410 /* pWLoop is a winner. Add it to the set of best so far */ 4411 pTo->maskLoop = pFrom->maskLoop | pWLoop->maskSelf; 4412 pTo->revLoop = revMask; 4413 pTo->nRow = nOut; 4414 pTo->rCost = rCost; 4415 pTo->rUnsorted = rUnsorted; 4416 pTo->isOrdered = isOrdered; 4417 memcpy(pTo->aLoop, pFrom->aLoop, sizeof(WhereLoop*)*iLoop); 4418 pTo->aLoop[iLoop] = pWLoop; 4419 if( nTo>=mxChoice ){ 4420 mxI = 0; 4421 mxCost = aTo[0].rCost; 4422 mxUnsorted = aTo[0].nRow; 4423 for(jj=1, pTo=&aTo[1]; jj<mxChoice; jj++, pTo++){ 4424 if( pTo->rCost>mxCost 4425 || (pTo->rCost==mxCost && pTo->rUnsorted>mxUnsorted) 4426 ){ 4427 mxCost = pTo->rCost; 4428 mxUnsorted = pTo->rUnsorted; 4429 mxI = jj; 4430 } 4431 } 4432 } 4433 } 4434 } 4435 4436 #ifdef WHERETRACE_ENABLED /* >=2 */ 4437 if( sqlite3WhereTrace & 0x02 ){ 4438 sqlite3DebugPrintf("---- after round %d ----\n", iLoop); 4439 for(ii=0, pTo=aTo; ii<nTo; ii++, pTo++){ 4440 sqlite3DebugPrintf(" %s cost=%-3d nrow=%-3d order=%c", 4441 wherePathName(pTo, iLoop+1, 0), pTo->rCost, pTo->nRow, 4442 pTo->isOrdered>=0 ? (pTo->isOrdered+'0') : '?'); 4443 if( pTo->isOrdered>0 ){ 4444 sqlite3DebugPrintf(" rev=0x%llx\n", pTo->revLoop); 4445 }else{ 4446 sqlite3DebugPrintf("\n"); 4447 } 4448 } 4449 } 4450 #endif 4451 4452 /* Swap the roles of aFrom and aTo for the next generation */ 4453 pFrom = aTo; 4454 aTo = aFrom; 4455 aFrom = pFrom; 4456 nFrom = nTo; 4457 } 4458 4459 if( nFrom==0 ){ 4460 sqlite3ErrorMsg(pParse, "no query solution"); 4461 sqlite3DbFreeNN(db, pSpace); 4462 return SQLITE_ERROR; 4463 } 4464 4465 /* Find the lowest cost path. pFrom will be left pointing to that path */ 4466 pFrom = aFrom; 4467 for(ii=1; ii<nFrom; ii++){ 4468 if( pFrom->rCost>aFrom[ii].rCost ) pFrom = &aFrom[ii]; 4469 } 4470 assert( pWInfo->nLevel==nLoop ); 4471 /* Load the lowest cost path into pWInfo */ 4472 for(iLoop=0; iLoop<nLoop; iLoop++){ 4473 WhereLevel *pLevel = pWInfo->a + iLoop; 4474 pLevel->pWLoop = pWLoop = pFrom->aLoop[iLoop]; 4475 pLevel->iFrom = pWLoop->iTab; 4476 pLevel->iTabCur = pWInfo->pTabList->a[pLevel->iFrom].iCursor; 4477 } 4478 if( (pWInfo->wctrlFlags & WHERE_WANT_DISTINCT)!=0 4479 && (pWInfo->wctrlFlags & WHERE_DISTINCTBY)==0 4480 && pWInfo->eDistinct==WHERE_DISTINCT_NOOP 4481 && nRowEst 4482 ){ 4483 Bitmask notUsed; 4484 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pResultSet, pFrom, 4485 WHERE_DISTINCTBY, nLoop-1, pFrom->aLoop[nLoop-1], ¬Used); 4486 if( rc==pWInfo->pResultSet->nExpr ){ 4487 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4488 } 4489 } 4490 pWInfo->bOrderedInnerLoop = 0; 4491 if( pWInfo->pOrderBy ){ 4492 if( pWInfo->wctrlFlags & WHERE_DISTINCTBY ){ 4493 if( pFrom->isOrdered==pWInfo->pOrderBy->nExpr ){ 4494 pWInfo->eDistinct = WHERE_DISTINCT_ORDERED; 4495 } 4496 }else{ 4497 pWInfo->nOBSat = pFrom->isOrdered; 4498 pWInfo->revMask = pFrom->revLoop; 4499 if( pWInfo->nOBSat<=0 ){ 4500 pWInfo->nOBSat = 0; 4501 if( nLoop>0 ){ 4502 u32 wsFlags = pFrom->aLoop[nLoop-1]->wsFlags; 4503 if( (wsFlags & WHERE_ONEROW)==0 4504 && (wsFlags&(WHERE_IPK|WHERE_COLUMN_IN))!=(WHERE_IPK|WHERE_COLUMN_IN) 4505 ){ 4506 Bitmask m = 0; 4507 int rc = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, pFrom, 4508 WHERE_ORDERBY_LIMIT, nLoop-1, pFrom->aLoop[nLoop-1], &m); 4509 testcase( wsFlags & WHERE_IPK ); 4510 testcase( wsFlags & WHERE_COLUMN_IN ); 4511 if( rc==pWInfo->pOrderBy->nExpr ){ 4512 pWInfo->bOrderedInnerLoop = 1; 4513 pWInfo->revMask = m; 4514 } 4515 } 4516 } 4517 }else if( nLoop 4518 && pWInfo->nOBSat==1 4519 && (pWInfo->wctrlFlags & (WHERE_ORDERBY_MIN|WHERE_ORDERBY_MAX))!=0 4520 ){ 4521 pWInfo->bOrderedInnerLoop = 1; 4522 } 4523 } 4524 if( (pWInfo->wctrlFlags & WHERE_SORTBYGROUP) 4525 && pWInfo->nOBSat==pWInfo->pOrderBy->nExpr && nLoop>0 4526 ){ 4527 Bitmask revMask = 0; 4528 int nOrder = wherePathSatisfiesOrderBy(pWInfo, pWInfo->pOrderBy, 4529 pFrom, 0, nLoop-1, pFrom->aLoop[nLoop-1], &revMask 4530 ); 4531 assert( pWInfo->sorted==0 ); 4532 if( nOrder==pWInfo->pOrderBy->nExpr ){ 4533 pWInfo->sorted = 1; 4534 pWInfo->revMask = revMask; 4535 } 4536 } 4537 } 4538 4539 4540 pWInfo->nRowOut = pFrom->nRow; 4541 4542 /* Free temporary memory and return success */ 4543 sqlite3DbFreeNN(db, pSpace); 4544 return SQLITE_OK; 4545 } 4546 4547 /* 4548 ** Most queries use only a single table (they are not joins) and have 4549 ** simple == constraints against indexed fields. This routine attempts 4550 ** to plan those simple cases using much less ceremony than the 4551 ** general-purpose query planner, and thereby yield faster sqlite3_prepare() 4552 ** times for the common case. 4553 ** 4554 ** Return non-zero on success, if this query can be handled by this 4555 ** no-frills query planner. Return zero if this query needs the 4556 ** general-purpose query planner. 4557 */ 4558 static int whereShortCut(WhereLoopBuilder *pBuilder){ 4559 WhereInfo *pWInfo; 4560 SrcItem *pItem; 4561 WhereClause *pWC; 4562 WhereTerm *pTerm; 4563 WhereLoop *pLoop; 4564 int iCur; 4565 int j; 4566 Table *pTab; 4567 Index *pIdx; 4568 4569 pWInfo = pBuilder->pWInfo; 4570 if( pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE ) return 0; 4571 assert( pWInfo->pTabList->nSrc>=1 ); 4572 pItem = pWInfo->pTabList->a; 4573 pTab = pItem->pTab; 4574 if( IsVirtual(pTab) ) return 0; 4575 if( pItem->fg.isIndexedBy ) return 0; 4576 iCur = pItem->iCursor; 4577 pWC = &pWInfo->sWC; 4578 pLoop = pBuilder->pNew; 4579 pLoop->wsFlags = 0; 4580 pLoop->nSkip = 0; 4581 pTerm = sqlite3WhereFindTerm(pWC, iCur, -1, 0, WO_EQ|WO_IS, 0); 4582 if( pTerm ){ 4583 testcase( pTerm->eOperator & WO_IS ); 4584 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_IPK|WHERE_ONEROW; 4585 pLoop->aLTerm[0] = pTerm; 4586 pLoop->nLTerm = 1; 4587 pLoop->u.btree.nEq = 1; 4588 /* TUNING: Cost of a rowid lookup is 10 */ 4589 pLoop->rRun = 33; /* 33==sqlite3LogEst(10) */ 4590 }else{ 4591 for(pIdx=pTab->pIndex; pIdx; pIdx=pIdx->pNext){ 4592 int opMask; 4593 assert( pLoop->aLTermSpace==pLoop->aLTerm ); 4594 if( !IsUniqueIndex(pIdx) 4595 || pIdx->pPartIdxWhere!=0 4596 || pIdx->nKeyCol>ArraySize(pLoop->aLTermSpace) 4597 ) continue; 4598 opMask = pIdx->uniqNotNull ? (WO_EQ|WO_IS) : WO_EQ; 4599 for(j=0; j<pIdx->nKeyCol; j++){ 4600 pTerm = sqlite3WhereFindTerm(pWC, iCur, j, 0, opMask, pIdx); 4601 if( pTerm==0 ) break; 4602 testcase( pTerm->eOperator & WO_IS ); 4603 pLoop->aLTerm[j] = pTerm; 4604 } 4605 if( j!=pIdx->nKeyCol ) continue; 4606 pLoop->wsFlags = WHERE_COLUMN_EQ|WHERE_ONEROW|WHERE_INDEXED; 4607 if( pIdx->isCovering || (pItem->colUsed & pIdx->colNotIdxed)==0 ){ 4608 pLoop->wsFlags |= WHERE_IDX_ONLY; 4609 } 4610 pLoop->nLTerm = j; 4611 pLoop->u.btree.nEq = j; 4612 pLoop->u.btree.pIndex = pIdx; 4613 /* TUNING: Cost of a unique index lookup is 15 */ 4614 pLoop->rRun = 39; /* 39==sqlite3LogEst(15) */ 4615 break; 4616 } 4617 } 4618 if( pLoop->wsFlags ){ 4619 pLoop->nOut = (LogEst)1; 4620 pWInfo->a[0].pWLoop = pLoop; 4621 assert( pWInfo->sMaskSet.n==1 && iCur==pWInfo->sMaskSet.ix[0] ); 4622 pLoop->maskSelf = 1; /* sqlite3WhereGetMask(&pWInfo->sMaskSet, iCur); */ 4623 pWInfo->a[0].iTabCur = iCur; 4624 pWInfo->nRowOut = 1; 4625 if( pWInfo->pOrderBy ) pWInfo->nOBSat = pWInfo->pOrderBy->nExpr; 4626 if( pWInfo->wctrlFlags & WHERE_WANT_DISTINCT ){ 4627 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4628 } 4629 #ifdef SQLITE_DEBUG 4630 pLoop->cId = '0'; 4631 #endif 4632 return 1; 4633 } 4634 return 0; 4635 } 4636 4637 /* 4638 ** Helper function for exprIsDeterministic(). 4639 */ 4640 static int exprNodeIsDeterministic(Walker *pWalker, Expr *pExpr){ 4641 if( pExpr->op==TK_FUNCTION && ExprHasProperty(pExpr, EP_ConstFunc)==0 ){ 4642 pWalker->eCode = 0; 4643 return WRC_Abort; 4644 } 4645 return WRC_Continue; 4646 } 4647 4648 /* 4649 ** Return true if the expression contains no non-deterministic SQL 4650 ** functions. Do not consider non-deterministic SQL functions that are 4651 ** part of sub-select statements. 4652 */ 4653 static int exprIsDeterministic(Expr *p){ 4654 Walker w; 4655 memset(&w, 0, sizeof(w)); 4656 w.eCode = 1; 4657 w.xExprCallback = exprNodeIsDeterministic; 4658 w.xSelectCallback = sqlite3SelectWalkFail; 4659 sqlite3WalkExpr(&w, p); 4660 return w.eCode; 4661 } 4662 4663 4664 #ifdef WHERETRACE_ENABLED 4665 /* 4666 ** Display all WhereLoops in pWInfo 4667 */ 4668 static void showAllWhereLoops(WhereInfo *pWInfo, WhereClause *pWC){ 4669 if( sqlite3WhereTrace ){ /* Display all of the WhereLoop objects */ 4670 WhereLoop *p; 4671 int i; 4672 static const char zLabel[] = "0123456789abcdefghijklmnopqrstuvwyxz" 4673 "ABCDEFGHIJKLMNOPQRSTUVWYXZ"; 4674 for(p=pWInfo->pLoops, i=0; p; p=p->pNextLoop, i++){ 4675 p->cId = zLabel[i%(sizeof(zLabel)-1)]; 4676 sqlite3WhereLoopPrint(p, pWC); 4677 } 4678 } 4679 } 4680 # define WHERETRACE_ALL_LOOPS(W,C) showAllWhereLoops(W,C) 4681 #else 4682 # define WHERETRACE_ALL_LOOPS(W,C) 4683 #endif 4684 4685 /* 4686 ** Generate the beginning of the loop used for WHERE clause processing. 4687 ** The return value is a pointer to an opaque structure that contains 4688 ** information needed to terminate the loop. Later, the calling routine 4689 ** should invoke sqlite3WhereEnd() with the return value of this function 4690 ** in order to complete the WHERE clause processing. 4691 ** 4692 ** If an error occurs, this routine returns NULL. 4693 ** 4694 ** The basic idea is to do a nested loop, one loop for each table in 4695 ** the FROM clause of a select. (INSERT and UPDATE statements are the 4696 ** same as a SELECT with only a single table in the FROM clause.) For 4697 ** example, if the SQL is this: 4698 ** 4699 ** SELECT * FROM t1, t2, t3 WHERE ...; 4700 ** 4701 ** Then the code generated is conceptually like the following: 4702 ** 4703 ** foreach row1 in t1 do \ Code generated 4704 ** foreach row2 in t2 do |-- by sqlite3WhereBegin() 4705 ** foreach row3 in t3 do / 4706 ** ... 4707 ** end \ Code generated 4708 ** end |-- by sqlite3WhereEnd() 4709 ** end / 4710 ** 4711 ** Note that the loops might not be nested in the order in which they 4712 ** appear in the FROM clause if a different order is better able to make 4713 ** use of indices. Note also that when the IN operator appears in 4714 ** the WHERE clause, it might result in additional nested loops for 4715 ** scanning through all values on the right-hand side of the IN. 4716 ** 4717 ** There are Btree cursors associated with each table. t1 uses cursor 4718 ** number pTabList->a[0].iCursor. t2 uses the cursor pTabList->a[1].iCursor. 4719 ** And so forth. This routine generates code to open those VDBE cursors 4720 ** and sqlite3WhereEnd() generates the code to close them. 4721 ** 4722 ** The code that sqlite3WhereBegin() generates leaves the cursors named 4723 ** in pTabList pointing at their appropriate entries. The [...] code 4724 ** can use OP_Column and OP_Rowid opcodes on these cursors to extract 4725 ** data from the various tables of the loop. 4726 ** 4727 ** If the WHERE clause is empty, the foreach loops must each scan their 4728 ** entire tables. Thus a three-way join is an O(N^3) operation. But if 4729 ** the tables have indices and there are terms in the WHERE clause that 4730 ** refer to those indices, a complete table scan can be avoided and the 4731 ** code will run much faster. Most of the work of this routine is checking 4732 ** to see if there are indices that can be used to speed up the loop. 4733 ** 4734 ** Terms of the WHERE clause are also used to limit which rows actually 4735 ** make it to the "..." in the middle of the loop. After each "foreach", 4736 ** terms of the WHERE clause that use only terms in that loop and outer 4737 ** loops are evaluated and if false a jump is made around all subsequent 4738 ** inner loops (or around the "..." if the test occurs within the inner- 4739 ** most loop) 4740 ** 4741 ** OUTER JOINS 4742 ** 4743 ** An outer join of tables t1 and t2 is conceptally coded as follows: 4744 ** 4745 ** foreach row1 in t1 do 4746 ** flag = 0 4747 ** foreach row2 in t2 do 4748 ** start: 4749 ** ... 4750 ** flag = 1 4751 ** end 4752 ** if flag==0 then 4753 ** move the row2 cursor to a null row 4754 ** goto start 4755 ** fi 4756 ** end 4757 ** 4758 ** ORDER BY CLAUSE PROCESSING 4759 ** 4760 ** pOrderBy is a pointer to the ORDER BY clause (or the GROUP BY clause 4761 ** if the WHERE_GROUPBY flag is set in wctrlFlags) of a SELECT statement 4762 ** if there is one. If there is no ORDER BY clause or if this routine 4763 ** is called from an UPDATE or DELETE statement, then pOrderBy is NULL. 4764 ** 4765 ** The iIdxCur parameter is the cursor number of an index. If 4766 ** WHERE_OR_SUBCLAUSE is set, iIdxCur is the cursor number of an index 4767 ** to use for OR clause processing. The WHERE clause should use this 4768 ** specific cursor. If WHERE_ONEPASS_DESIRED is set, then iIdxCur is 4769 ** the first cursor in an array of cursors for all indices. iIdxCur should 4770 ** be used to compute the appropriate cursor depending on which index is 4771 ** used. 4772 */ 4773 WhereInfo *sqlite3WhereBegin( 4774 Parse *pParse, /* The parser context */ 4775 SrcList *pTabList, /* FROM clause: A list of all tables to be scanned */ 4776 Expr *pWhere, /* The WHERE clause */ 4777 ExprList *pOrderBy, /* An ORDER BY (or GROUP BY) clause, or NULL */ 4778 ExprList *pResultSet, /* Query result set. Req'd for DISTINCT */ 4779 u16 wctrlFlags, /* The WHERE_* flags defined in sqliteInt.h */ 4780 int iAuxArg /* If WHERE_OR_SUBCLAUSE is set, index cursor number 4781 ** If WHERE_USE_LIMIT, then the limit amount */ 4782 ){ 4783 int nByteWInfo; /* Num. bytes allocated for WhereInfo struct */ 4784 int nTabList; /* Number of elements in pTabList */ 4785 WhereInfo *pWInfo; /* Will become the return value of this function */ 4786 Vdbe *v = pParse->pVdbe; /* The virtual database engine */ 4787 Bitmask notReady; /* Cursors that are not yet positioned */ 4788 WhereLoopBuilder sWLB; /* The WhereLoop builder */ 4789 WhereMaskSet *pMaskSet; /* The expression mask set */ 4790 WhereLevel *pLevel; /* A single level in pWInfo->a[] */ 4791 WhereLoop *pLoop; /* Pointer to a single WhereLoop object */ 4792 int ii; /* Loop counter */ 4793 sqlite3 *db; /* Database connection */ 4794 int rc; /* Return code */ 4795 u8 bFordelete = 0; /* OPFLAG_FORDELETE or zero, as appropriate */ 4796 4797 assert( (wctrlFlags & WHERE_ONEPASS_MULTIROW)==0 || ( 4798 (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 4799 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4800 )); 4801 4802 /* Only one of WHERE_OR_SUBCLAUSE or WHERE_USE_LIMIT */ 4803 assert( (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 4804 || (wctrlFlags & WHERE_USE_LIMIT)==0 ); 4805 4806 /* Variable initialization */ 4807 db = pParse->db; 4808 memset(&sWLB, 0, sizeof(sWLB)); 4809 4810 /* An ORDER/GROUP BY clause of more than 63 terms cannot be optimized */ 4811 testcase( pOrderBy && pOrderBy->nExpr==BMS-1 ); 4812 if( pOrderBy && pOrderBy->nExpr>=BMS ) pOrderBy = 0; 4813 sWLB.pOrderBy = pOrderBy; 4814 4815 /* Disable the DISTINCT optimization if SQLITE_DistinctOpt is set via 4816 ** sqlite3_test_ctrl(SQLITE_TESTCTRL_OPTIMIZATIONS,...) */ 4817 if( OptimizationDisabled(db, SQLITE_DistinctOpt) ){ 4818 wctrlFlags &= ~WHERE_WANT_DISTINCT; 4819 } 4820 4821 /* The number of tables in the FROM clause is limited by the number of 4822 ** bits in a Bitmask 4823 */ 4824 testcase( pTabList->nSrc==BMS ); 4825 if( pTabList->nSrc>BMS ){ 4826 sqlite3ErrorMsg(pParse, "at most %d tables in a join", BMS); 4827 return 0; 4828 } 4829 4830 /* This function normally generates a nested loop for all tables in 4831 ** pTabList. But if the WHERE_OR_SUBCLAUSE flag is set, then we should 4832 ** only generate code for the first table in pTabList and assume that 4833 ** any cursors associated with subsequent tables are uninitialized. 4834 */ 4835 nTabList = (wctrlFlags & WHERE_OR_SUBCLAUSE) ? 1 : pTabList->nSrc; 4836 4837 /* Allocate and initialize the WhereInfo structure that will become the 4838 ** return value. A single allocation is used to store the WhereInfo 4839 ** struct, the contents of WhereInfo.a[], the WhereClause structure 4840 ** and the WhereMaskSet structure. Since WhereClause contains an 8-byte 4841 ** field (type Bitmask) it must be aligned on an 8-byte boundary on 4842 ** some architectures. Hence the ROUND8() below. 4843 */ 4844 nByteWInfo = ROUND8(sizeof(WhereInfo)+(nTabList-1)*sizeof(WhereLevel)); 4845 pWInfo = sqlite3DbMallocRawNN(db, nByteWInfo + sizeof(WhereLoop)); 4846 if( db->mallocFailed ){ 4847 sqlite3DbFree(db, pWInfo); 4848 pWInfo = 0; 4849 goto whereBeginError; 4850 } 4851 pWInfo->pParse = pParse; 4852 pWInfo->pTabList = pTabList; 4853 pWInfo->pOrderBy = pOrderBy; 4854 pWInfo->pWhere = pWhere; 4855 pWInfo->pResultSet = pResultSet; 4856 pWInfo->aiCurOnePass[0] = pWInfo->aiCurOnePass[1] = -1; 4857 pWInfo->nLevel = nTabList; 4858 pWInfo->iBreak = pWInfo->iContinue = sqlite3VdbeMakeLabel(pParse); 4859 pWInfo->wctrlFlags = wctrlFlags; 4860 pWInfo->iLimit = iAuxArg; 4861 pWInfo->savedNQueryLoop = pParse->nQueryLoop; 4862 memset(&pWInfo->nOBSat, 0, 4863 offsetof(WhereInfo,sWC) - offsetof(WhereInfo,nOBSat)); 4864 memset(&pWInfo->a[0], 0, sizeof(WhereLoop)+nTabList*sizeof(WhereLevel)); 4865 assert( pWInfo->eOnePass==ONEPASS_OFF ); /* ONEPASS defaults to OFF */ 4866 pMaskSet = &pWInfo->sMaskSet; 4867 sWLB.pWInfo = pWInfo; 4868 sWLB.pWC = &pWInfo->sWC; 4869 sWLB.pNew = (WhereLoop*)(((char*)pWInfo)+nByteWInfo); 4870 assert( EIGHT_BYTE_ALIGNMENT(sWLB.pNew) ); 4871 whereLoopInit(sWLB.pNew); 4872 #ifdef SQLITE_DEBUG 4873 sWLB.pNew->cId = '*'; 4874 #endif 4875 4876 /* Split the WHERE clause into separate subexpressions where each 4877 ** subexpression is separated by an AND operator. 4878 */ 4879 initMaskSet(pMaskSet); 4880 sqlite3WhereClauseInit(&pWInfo->sWC, pWInfo); 4881 sqlite3WhereSplit(&pWInfo->sWC, pWhere, TK_AND); 4882 4883 /* Special case: No FROM clause 4884 */ 4885 if( nTabList==0 ){ 4886 if( pOrderBy ) pWInfo->nOBSat = pOrderBy->nExpr; 4887 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4888 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4889 } 4890 ExplainQueryPlan((pParse, 0, "SCAN CONSTANT ROW")); 4891 }else{ 4892 /* Assign a bit from the bitmask to every term in the FROM clause. 4893 ** 4894 ** The N-th term of the FROM clause is assigned a bitmask of 1<<N. 4895 ** 4896 ** The rule of the previous sentence ensures thta if X is the bitmask for 4897 ** a table T, then X-1 is the bitmask for all other tables to the left of T. 4898 ** Knowing the bitmask for all tables to the left of a left join is 4899 ** important. Ticket #3015. 4900 ** 4901 ** Note that bitmasks are created for all pTabList->nSrc tables in 4902 ** pTabList, not just the first nTabList tables. nTabList is normally 4903 ** equal to pTabList->nSrc but might be shortened to 1 if the 4904 ** WHERE_OR_SUBCLAUSE flag is set. 4905 */ 4906 ii = 0; 4907 do{ 4908 createMask(pMaskSet, pTabList->a[ii].iCursor); 4909 sqlite3WhereTabFuncArgs(pParse, &pTabList->a[ii], &pWInfo->sWC); 4910 }while( (++ii)<pTabList->nSrc ); 4911 #ifdef SQLITE_DEBUG 4912 { 4913 Bitmask mx = 0; 4914 for(ii=0; ii<pTabList->nSrc; ii++){ 4915 Bitmask m = sqlite3WhereGetMask(pMaskSet, pTabList->a[ii].iCursor); 4916 assert( m>=mx ); 4917 mx = m; 4918 } 4919 } 4920 #endif 4921 } 4922 4923 /* Analyze all of the subexpressions. */ 4924 sqlite3WhereExprAnalyze(pTabList, &pWInfo->sWC); 4925 if( db->mallocFailed ) goto whereBeginError; 4926 4927 /* Special case: WHERE terms that do not refer to any tables in the join 4928 ** (constant expressions). Evaluate each such term, and jump over all the 4929 ** generated code if the result is not true. 4930 ** 4931 ** Do not do this if the expression contains non-deterministic functions 4932 ** that are not within a sub-select. This is not strictly required, but 4933 ** preserves SQLite's legacy behaviour in the following two cases: 4934 ** 4935 ** FROM ... WHERE random()>0; -- eval random() once per row 4936 ** FROM ... WHERE (SELECT random())>0; -- eval random() once overall 4937 */ 4938 for(ii=0; ii<sWLB.pWC->nTerm; ii++){ 4939 WhereTerm *pT = &sWLB.pWC->a[ii]; 4940 if( pT->wtFlags & TERM_VIRTUAL ) continue; 4941 if( pT->prereqAll==0 && (nTabList==0 || exprIsDeterministic(pT->pExpr)) ){ 4942 sqlite3ExprIfFalse(pParse, pT->pExpr, pWInfo->iBreak, SQLITE_JUMPIFNULL); 4943 pT->wtFlags |= TERM_CODED; 4944 } 4945 } 4946 4947 if( wctrlFlags & WHERE_WANT_DISTINCT ){ 4948 if( isDistinctRedundant(pParse, pTabList, &pWInfo->sWC, pResultSet) ){ 4949 /* The DISTINCT marking is pointless. Ignore it. */ 4950 pWInfo->eDistinct = WHERE_DISTINCT_UNIQUE; 4951 }else if( pOrderBy==0 ){ 4952 /* Try to ORDER BY the result set to make distinct processing easier */ 4953 pWInfo->wctrlFlags |= WHERE_DISTINCTBY; 4954 pWInfo->pOrderBy = pResultSet; 4955 } 4956 } 4957 4958 /* Construct the WhereLoop objects */ 4959 #if defined(WHERETRACE_ENABLED) 4960 if( sqlite3WhereTrace & 0xffff ){ 4961 sqlite3DebugPrintf("*** Optimizer Start *** (wctrlFlags: 0x%x",wctrlFlags); 4962 if( wctrlFlags & WHERE_USE_LIMIT ){ 4963 sqlite3DebugPrintf(", limit: %d", iAuxArg); 4964 } 4965 sqlite3DebugPrintf(")\n"); 4966 if( sqlite3WhereTrace & 0x100 ){ 4967 Select sSelect; 4968 memset(&sSelect, 0, sizeof(sSelect)); 4969 sSelect.selFlags = SF_WhereBegin; 4970 sSelect.pSrc = pTabList; 4971 sSelect.pWhere = pWhere; 4972 sSelect.pOrderBy = pOrderBy; 4973 sSelect.pEList = pResultSet; 4974 sqlite3TreeViewSelect(0, &sSelect, 0); 4975 } 4976 } 4977 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 4978 sqlite3DebugPrintf("---- WHERE clause at start of analysis:\n"); 4979 sqlite3WhereClausePrint(sWLB.pWC); 4980 } 4981 #endif 4982 4983 if( nTabList!=1 || whereShortCut(&sWLB)==0 ){ 4984 rc = whereLoopAddAll(&sWLB); 4985 if( rc ) goto whereBeginError; 4986 4987 #ifdef SQLITE_ENABLE_STAT4 4988 /* If one or more WhereTerm.truthProb values were used in estimating 4989 ** loop parameters, but then those truthProb values were subsequently 4990 ** changed based on STAT4 information while computing subsequent loops, 4991 ** then we need to rerun the whole loop building process so that all 4992 ** loops will be built using the revised truthProb values. */ 4993 if( sWLB.bldFlags2 & SQLITE_BLDF2_2NDPASS ){ 4994 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 4995 WHERETRACE(0xffff, 4996 ("**** Redo all loop computations due to" 4997 " TERM_HIGHTRUTH changes ****\n")); 4998 while( pWInfo->pLoops ){ 4999 WhereLoop *p = pWInfo->pLoops; 5000 pWInfo->pLoops = p->pNextLoop; 5001 whereLoopDelete(db, p); 5002 } 5003 rc = whereLoopAddAll(&sWLB); 5004 if( rc ) goto whereBeginError; 5005 } 5006 #endif 5007 WHERETRACE_ALL_LOOPS(pWInfo, sWLB.pWC); 5008 5009 wherePathSolver(pWInfo, 0); 5010 if( db->mallocFailed ) goto whereBeginError; 5011 if( pWInfo->pOrderBy ){ 5012 wherePathSolver(pWInfo, pWInfo->nRowOut+1); 5013 if( db->mallocFailed ) goto whereBeginError; 5014 } 5015 } 5016 if( pWInfo->pOrderBy==0 && (db->flags & SQLITE_ReverseOrder)!=0 ){ 5017 pWInfo->revMask = ALLBITS; 5018 } 5019 if( pParse->nErr || db->mallocFailed ){ 5020 goto whereBeginError; 5021 } 5022 #ifdef WHERETRACE_ENABLED 5023 if( sqlite3WhereTrace ){ 5024 sqlite3DebugPrintf("---- Solution nRow=%d", pWInfo->nRowOut); 5025 if( pWInfo->nOBSat>0 ){ 5026 sqlite3DebugPrintf(" ORDERBY=%d,0x%llx", pWInfo->nOBSat, pWInfo->revMask); 5027 } 5028 switch( pWInfo->eDistinct ){ 5029 case WHERE_DISTINCT_UNIQUE: { 5030 sqlite3DebugPrintf(" DISTINCT=unique"); 5031 break; 5032 } 5033 case WHERE_DISTINCT_ORDERED: { 5034 sqlite3DebugPrintf(" DISTINCT=ordered"); 5035 break; 5036 } 5037 case WHERE_DISTINCT_UNORDERED: { 5038 sqlite3DebugPrintf(" DISTINCT=unordered"); 5039 break; 5040 } 5041 } 5042 sqlite3DebugPrintf("\n"); 5043 for(ii=0; ii<pWInfo->nLevel; ii++){ 5044 sqlite3WhereLoopPrint(pWInfo->a[ii].pWLoop, sWLB.pWC); 5045 } 5046 } 5047 #endif 5048 5049 /* Attempt to omit tables from the join that do not affect the result. 5050 ** For a table to not affect the result, the following must be true: 5051 ** 5052 ** 1) The query must not be an aggregate. 5053 ** 2) The table must be the RHS of a LEFT JOIN. 5054 ** 3) Either the query must be DISTINCT, or else the ON or USING clause 5055 ** must contain a constraint that limits the scan of the table to 5056 ** at most a single row. 5057 ** 4) The table must not be referenced by any part of the query apart 5058 ** from its own USING or ON clause. 5059 ** 5060 ** For example, given: 5061 ** 5062 ** CREATE TABLE t1(ipk INTEGER PRIMARY KEY, v1); 5063 ** CREATE TABLE t2(ipk INTEGER PRIMARY KEY, v2); 5064 ** CREATE TABLE t3(ipk INTEGER PRIMARY KEY, v3); 5065 ** 5066 ** then table t2 can be omitted from the following: 5067 ** 5068 ** SELECT v1, v3 FROM t1 5069 ** LEFT JOIN t2 ON (t1.ipk=t2.ipk) 5070 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 5071 ** 5072 ** or from: 5073 ** 5074 ** SELECT DISTINCT v1, v3 FROM t1 5075 ** LEFT JOIN t2 5076 ** LEFT JOIN t3 ON (t1.ipk=t3.ipk) 5077 */ 5078 notReady = ~(Bitmask)0; 5079 if( pWInfo->nLevel>=2 5080 && pResultSet!=0 /* these two combine to guarantee */ 5081 && 0==(wctrlFlags & WHERE_AGG_DISTINCT) /* condition (1) above */ 5082 && OptimizationEnabled(db, SQLITE_OmitNoopJoin) 5083 ){ 5084 int i; 5085 Bitmask tabUsed = sqlite3WhereExprListUsage(pMaskSet, pResultSet); 5086 if( sWLB.pOrderBy ){ 5087 tabUsed |= sqlite3WhereExprListUsage(pMaskSet, sWLB.pOrderBy); 5088 } 5089 for(i=pWInfo->nLevel-1; i>=1; i--){ 5090 WhereTerm *pTerm, *pEnd; 5091 SrcItem *pItem; 5092 pLoop = pWInfo->a[i].pWLoop; 5093 pItem = &pWInfo->pTabList->a[pLoop->iTab]; 5094 if( (pItem->fg.jointype & JT_LEFT)==0 ) continue; 5095 if( (wctrlFlags & WHERE_WANT_DISTINCT)==0 5096 && (pLoop->wsFlags & WHERE_ONEROW)==0 5097 ){ 5098 continue; 5099 } 5100 if( (tabUsed & pLoop->maskSelf)!=0 ) continue; 5101 pEnd = sWLB.pWC->a + sWLB.pWC->nTerm; 5102 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 5103 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5104 if( !ExprHasProperty(pTerm->pExpr, EP_FromJoin) 5105 || pTerm->pExpr->iRightJoinTable!=pItem->iCursor 5106 ){ 5107 break; 5108 } 5109 } 5110 } 5111 if( pTerm<pEnd ) continue; 5112 WHERETRACE(0xffff, ("-> drop loop %c not used\n", pLoop->cId)); 5113 notReady &= ~pLoop->maskSelf; 5114 for(pTerm=sWLB.pWC->a; pTerm<pEnd; pTerm++){ 5115 if( (pTerm->prereqAll & pLoop->maskSelf)!=0 ){ 5116 pTerm->wtFlags |= TERM_CODED; 5117 } 5118 } 5119 if( i!=pWInfo->nLevel-1 ){ 5120 int nByte = (pWInfo->nLevel-1-i) * sizeof(WhereLevel); 5121 memmove(&pWInfo->a[i], &pWInfo->a[i+1], nByte); 5122 } 5123 pWInfo->nLevel--; 5124 nTabList--; 5125 } 5126 } 5127 #if defined(WHERETRACE_ENABLED) 5128 if( sqlite3WhereTrace & 0x100 ){ /* Display all terms of the WHERE clause */ 5129 sqlite3DebugPrintf("---- WHERE clause at end of analysis:\n"); 5130 sqlite3WhereClausePrint(sWLB.pWC); 5131 } 5132 WHERETRACE(0xffff,("*** Optimizer Finished ***\n")); 5133 #endif 5134 pWInfo->pParse->nQueryLoop += pWInfo->nRowOut; 5135 5136 /* If the caller is an UPDATE or DELETE statement that is requesting 5137 ** to use a one-pass algorithm, determine if this is appropriate. 5138 ** 5139 ** A one-pass approach can be used if the caller has requested one 5140 ** and either (a) the scan visits at most one row or (b) each 5141 ** of the following are true: 5142 ** 5143 ** * the caller has indicated that a one-pass approach can be used 5144 ** with multiple rows (by setting WHERE_ONEPASS_MULTIROW), and 5145 ** * the table is not a virtual table, and 5146 ** * either the scan does not use the OR optimization or the caller 5147 ** is a DELETE operation (WHERE_DUPLICATES_OK is only specified 5148 ** for DELETE). 5149 ** 5150 ** The last qualification is because an UPDATE statement uses 5151 ** WhereInfo.aiCurOnePass[1] to determine whether or not it really can 5152 ** use a one-pass approach, and this is not set accurately for scans 5153 ** that use the OR optimization. 5154 */ 5155 assert( (wctrlFlags & WHERE_ONEPASS_DESIRED)==0 || pWInfo->nLevel==1 ); 5156 if( (wctrlFlags & WHERE_ONEPASS_DESIRED)!=0 ){ 5157 int wsFlags = pWInfo->a[0].pWLoop->wsFlags; 5158 int bOnerow = (wsFlags & WHERE_ONEROW)!=0; 5159 assert( !(wsFlags & WHERE_VIRTUALTABLE) || IsVirtual(pTabList->a[0].pTab) ); 5160 if( bOnerow || ( 5161 0!=(wctrlFlags & WHERE_ONEPASS_MULTIROW) 5162 && !IsVirtual(pTabList->a[0].pTab) 5163 && (0==(wsFlags & WHERE_MULTI_OR) || (wctrlFlags & WHERE_DUPLICATES_OK)) 5164 )){ 5165 pWInfo->eOnePass = bOnerow ? ONEPASS_SINGLE : ONEPASS_MULTI; 5166 if( HasRowid(pTabList->a[0].pTab) && (wsFlags & WHERE_IDX_ONLY) ){ 5167 if( wctrlFlags & WHERE_ONEPASS_MULTIROW ){ 5168 bFordelete = OPFLAG_FORDELETE; 5169 } 5170 pWInfo->a[0].pWLoop->wsFlags = (wsFlags & ~WHERE_IDX_ONLY); 5171 } 5172 } 5173 } 5174 5175 /* Open all tables in the pTabList and any indices selected for 5176 ** searching those tables. 5177 */ 5178 for(ii=0, pLevel=pWInfo->a; ii<nTabList; ii++, pLevel++){ 5179 Table *pTab; /* Table to open */ 5180 int iDb; /* Index of database containing table/index */ 5181 SrcItem *pTabItem; 5182 5183 pTabItem = &pTabList->a[pLevel->iFrom]; 5184 pTab = pTabItem->pTab; 5185 iDb = sqlite3SchemaToIndex(db, pTab->pSchema); 5186 pLoop = pLevel->pWLoop; 5187 if( (pTab->tabFlags & TF_Ephemeral)!=0 || pTab->pSelect ){ 5188 /* Do nothing */ 5189 }else 5190 #ifndef SQLITE_OMIT_VIRTUALTABLE 5191 if( (pLoop->wsFlags & WHERE_VIRTUALTABLE)!=0 ){ 5192 const char *pVTab = (const char *)sqlite3GetVTable(db, pTab); 5193 int iCur = pTabItem->iCursor; 5194 sqlite3VdbeAddOp4(v, OP_VOpen, iCur, 0, 0, pVTab, P4_VTAB); 5195 }else if( IsVirtual(pTab) ){ 5196 /* noop */ 5197 }else 5198 #endif 5199 if( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 5200 && (wctrlFlags & WHERE_OR_SUBCLAUSE)==0 ){ 5201 int op = OP_OpenRead; 5202 if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5203 op = OP_OpenWrite; 5204 pWInfo->aiCurOnePass[0] = pTabItem->iCursor; 5205 }; 5206 sqlite3OpenTable(pParse, pTabItem->iCursor, iDb, pTab, op); 5207 assert( pTabItem->iCursor==pLevel->iTabCur ); 5208 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS-1 ); 5209 testcase( pWInfo->eOnePass==ONEPASS_OFF && pTab->nCol==BMS ); 5210 if( pWInfo->eOnePass==ONEPASS_OFF 5211 && pTab->nCol<BMS 5212 && (pTab->tabFlags & (TF_HasGenerated|TF_WithoutRowid))==0 5213 ){ 5214 /* If we know that only a prefix of the record will be used, 5215 ** it is advantageous to reduce the "column count" field in 5216 ** the P4 operand of the OP_OpenRead/Write opcode. */ 5217 Bitmask b = pTabItem->colUsed; 5218 int n = 0; 5219 for(; b; b=b>>1, n++){} 5220 sqlite3VdbeChangeP4(v, -1, SQLITE_INT_TO_PTR(n), P4_INT32); 5221 assert( n<=pTab->nCol ); 5222 } 5223 #ifdef SQLITE_ENABLE_CURSOR_HINTS 5224 if( pLoop->u.btree.pIndex!=0 ){ 5225 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ|bFordelete); 5226 }else 5227 #endif 5228 { 5229 sqlite3VdbeChangeP5(v, bFordelete); 5230 } 5231 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5232 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, pTabItem->iCursor, 0, 0, 5233 (const u8*)&pTabItem->colUsed, P4_INT64); 5234 #endif 5235 }else{ 5236 sqlite3TableLock(pParse, iDb, pTab->tnum, 0, pTab->zName); 5237 } 5238 if( pLoop->wsFlags & WHERE_INDEXED ){ 5239 Index *pIx = pLoop->u.btree.pIndex; 5240 int iIndexCur; 5241 int op = OP_OpenRead; 5242 /* iAuxArg is always set to a positive value if ONEPASS is possible */ 5243 assert( iAuxArg!=0 || (pWInfo->wctrlFlags & WHERE_ONEPASS_DESIRED)==0 ); 5244 if( !HasRowid(pTab) && IsPrimaryKeyIndex(pIx) 5245 && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 5246 ){ 5247 /* This is one term of an OR-optimization using the PRIMARY KEY of a 5248 ** WITHOUT ROWID table. No need for a separate index */ 5249 iIndexCur = pLevel->iTabCur; 5250 op = 0; 5251 }else if( pWInfo->eOnePass!=ONEPASS_OFF ){ 5252 Index *pJ = pTabItem->pTab->pIndex; 5253 iIndexCur = iAuxArg; 5254 assert( wctrlFlags & WHERE_ONEPASS_DESIRED ); 5255 while( ALWAYS(pJ) && pJ!=pIx ){ 5256 iIndexCur++; 5257 pJ = pJ->pNext; 5258 } 5259 op = OP_OpenWrite; 5260 pWInfo->aiCurOnePass[1] = iIndexCur; 5261 }else if( iAuxArg && (wctrlFlags & WHERE_OR_SUBCLAUSE)!=0 ){ 5262 iIndexCur = iAuxArg; 5263 op = OP_ReopenIdx; 5264 }else{ 5265 iIndexCur = pParse->nTab++; 5266 } 5267 pLevel->iIdxCur = iIndexCur; 5268 assert( pIx->pSchema==pTab->pSchema ); 5269 assert( iIndexCur>=0 ); 5270 if( op ){ 5271 sqlite3VdbeAddOp3(v, op, iIndexCur, pIx->tnum, iDb); 5272 sqlite3VdbeSetP4KeyInfo(pParse, pIx); 5273 if( (pLoop->wsFlags & WHERE_CONSTRAINT)!=0 5274 && (pLoop->wsFlags & (WHERE_COLUMN_RANGE|WHERE_SKIPSCAN))==0 5275 && (pLoop->wsFlags & WHERE_BIGNULL_SORT)==0 5276 && (pLoop->wsFlags & WHERE_IN_SEEKSCAN)==0 5277 && (pWInfo->wctrlFlags&WHERE_ORDERBY_MIN)==0 5278 && pWInfo->eDistinct!=WHERE_DISTINCT_ORDERED 5279 ){ 5280 sqlite3VdbeChangeP5(v, OPFLAG_SEEKEQ); 5281 } 5282 VdbeComment((v, "%s", pIx->zName)); 5283 #ifdef SQLITE_ENABLE_COLUMN_USED_MASK 5284 { 5285 u64 colUsed = 0; 5286 int ii, jj; 5287 for(ii=0; ii<pIx->nColumn; ii++){ 5288 jj = pIx->aiColumn[ii]; 5289 if( jj<0 ) continue; 5290 if( jj>63 ) jj = 63; 5291 if( (pTabItem->colUsed & MASKBIT(jj))==0 ) continue; 5292 colUsed |= ((u64)1)<<(ii<63 ? ii : 63); 5293 } 5294 sqlite3VdbeAddOp4Dup8(v, OP_ColumnsUsed, iIndexCur, 0, 0, 5295 (u8*)&colUsed, P4_INT64); 5296 } 5297 #endif /* SQLITE_ENABLE_COLUMN_USED_MASK */ 5298 } 5299 } 5300 if( iDb>=0 ) sqlite3CodeVerifySchema(pParse, iDb); 5301 } 5302 pWInfo->iTop = sqlite3VdbeCurrentAddr(v); 5303 if( db->mallocFailed ) goto whereBeginError; 5304 5305 /* Generate the code to do the search. Each iteration of the for 5306 ** loop below generates code for a single nested loop of the VM 5307 ** program. 5308 */ 5309 for(ii=0; ii<nTabList; ii++){ 5310 int addrExplain; 5311 int wsFlags; 5312 pLevel = &pWInfo->a[ii]; 5313 wsFlags = pLevel->pWLoop->wsFlags; 5314 #ifndef SQLITE_OMIT_AUTOMATIC_INDEX 5315 if( (pLevel->pWLoop->wsFlags & WHERE_AUTO_INDEX)!=0 ){ 5316 constructAutomaticIndex(pParse, &pWInfo->sWC, 5317 &pTabList->a[pLevel->iFrom], notReady, pLevel); 5318 if( db->mallocFailed ) goto whereBeginError; 5319 } 5320 #endif 5321 addrExplain = sqlite3WhereExplainOneScan( 5322 pParse, pTabList, pLevel, wctrlFlags 5323 ); 5324 pLevel->addrBody = sqlite3VdbeCurrentAddr(v); 5325 notReady = sqlite3WhereCodeOneLoopStart(pParse,v,pWInfo,ii,pLevel,notReady); 5326 pWInfo->iContinue = pLevel->addrCont; 5327 if( (wsFlags&WHERE_MULTI_OR)==0 && (wctrlFlags&WHERE_OR_SUBCLAUSE)==0 ){ 5328 sqlite3WhereAddScanStatus(v, pTabList, pLevel, addrExplain); 5329 } 5330 } 5331 5332 /* Done. */ 5333 VdbeModuleComment((v, "Begin WHERE-core")); 5334 pWInfo->iEndWhere = sqlite3VdbeCurrentAddr(v); 5335 return pWInfo; 5336 5337 /* Jump here if malloc fails */ 5338 whereBeginError: 5339 if( pWInfo ){ 5340 testcase( pWInfo->pExprMods!=0 ); 5341 whereUndoExprMods(pWInfo); 5342 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5343 whereInfoFree(db, pWInfo); 5344 } 5345 return 0; 5346 } 5347 5348 /* 5349 ** Part of sqlite3WhereEnd() will rewrite opcodes to reference the 5350 ** index rather than the main table. In SQLITE_DEBUG mode, we want 5351 ** to trace those changes if PRAGMA vdbe_addoptrace=on. This routine 5352 ** does that. 5353 */ 5354 #ifndef SQLITE_DEBUG 5355 # define OpcodeRewriteTrace(D,K,P) /* no-op */ 5356 #else 5357 # define OpcodeRewriteTrace(D,K,P) sqlite3WhereOpcodeRewriteTrace(D,K,P) 5358 static void sqlite3WhereOpcodeRewriteTrace( 5359 sqlite3 *db, 5360 int pc, 5361 VdbeOp *pOp 5362 ){ 5363 if( (db->flags & SQLITE_VdbeAddopTrace)==0 ) return; 5364 sqlite3VdbePrintOp(0, pc, pOp); 5365 } 5366 #endif 5367 5368 /* 5369 ** Generate the end of the WHERE loop. See comments on 5370 ** sqlite3WhereBegin() for additional information. 5371 */ 5372 void sqlite3WhereEnd(WhereInfo *pWInfo){ 5373 Parse *pParse = pWInfo->pParse; 5374 Vdbe *v = pParse->pVdbe; 5375 int i; 5376 WhereLevel *pLevel; 5377 WhereLoop *pLoop; 5378 SrcList *pTabList = pWInfo->pTabList; 5379 sqlite3 *db = pParse->db; 5380 int iEnd = sqlite3VdbeCurrentAddr(v); 5381 5382 /* Generate loop termination code. 5383 */ 5384 VdbeModuleComment((v, "End WHERE-core")); 5385 for(i=pWInfo->nLevel-1; i>=0; i--){ 5386 int addr; 5387 pLevel = &pWInfo->a[i]; 5388 pLoop = pLevel->pWLoop; 5389 if( pLevel->op!=OP_Noop ){ 5390 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5391 int addrSeek = 0; 5392 Index *pIdx; 5393 int n; 5394 if( pWInfo->eDistinct==WHERE_DISTINCT_ORDERED 5395 && i==pWInfo->nLevel-1 /* Ticket [ef9318757b152e3] 2017-10-21 */ 5396 && (pLoop->wsFlags & WHERE_INDEXED)!=0 5397 && (pIdx = pLoop->u.btree.pIndex)->hasStat1 5398 && (n = pLoop->u.btree.nDistinctCol)>0 5399 && pIdx->aiRowLogEst[n]>=36 5400 ){ 5401 int r1 = pParse->nMem+1; 5402 int j, op; 5403 for(j=0; j<n; j++){ 5404 sqlite3VdbeAddOp3(v, OP_Column, pLevel->iIdxCur, j, r1+j); 5405 } 5406 pParse->nMem += n+1; 5407 op = pLevel->op==OP_Prev ? OP_SeekLT : OP_SeekGT; 5408 addrSeek = sqlite3VdbeAddOp4Int(v, op, pLevel->iIdxCur, 0, r1, n); 5409 VdbeCoverageIf(v, op==OP_SeekLT); 5410 VdbeCoverageIf(v, op==OP_SeekGT); 5411 sqlite3VdbeAddOp2(v, OP_Goto, 1, pLevel->p2); 5412 } 5413 #endif /* SQLITE_DISABLE_SKIPAHEAD_DISTINCT */ 5414 /* The common case: Advance to the next row */ 5415 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5416 sqlite3VdbeAddOp3(v, pLevel->op, pLevel->p1, pLevel->p2, pLevel->p3); 5417 sqlite3VdbeChangeP5(v, pLevel->p5); 5418 VdbeCoverage(v); 5419 VdbeCoverageIf(v, pLevel->op==OP_Next); 5420 VdbeCoverageIf(v, pLevel->op==OP_Prev); 5421 VdbeCoverageIf(v, pLevel->op==OP_VNext); 5422 if( pLevel->regBignull ){ 5423 sqlite3VdbeResolveLabel(v, pLevel->addrBignull); 5424 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, pLevel->regBignull, pLevel->p2-1); 5425 VdbeCoverage(v); 5426 } 5427 #ifndef SQLITE_DISABLE_SKIPAHEAD_DISTINCT 5428 if( addrSeek ) sqlite3VdbeJumpHere(v, addrSeek); 5429 #endif 5430 }else{ 5431 sqlite3VdbeResolveLabel(v, pLevel->addrCont); 5432 } 5433 if( pLoop->wsFlags & WHERE_IN_ABLE && pLevel->u.in.nIn>0 ){ 5434 struct InLoop *pIn; 5435 int j; 5436 sqlite3VdbeResolveLabel(v, pLevel->addrNxt); 5437 for(j=pLevel->u.in.nIn, pIn=&pLevel->u.in.aInLoop[j-1]; j>0; j--, pIn--){ 5438 sqlite3VdbeJumpHere(v, pIn->addrInTop+1); 5439 if( pIn->eEndLoopOp!=OP_Noop ){ 5440 if( pIn->nPrefix ){ 5441 int bEarlyOut = 5442 (pLoop->wsFlags & WHERE_VIRTUALTABLE)==0 5443 && (pLoop->wsFlags & WHERE_IN_EARLYOUT)!=0; 5444 if( pLevel->iLeftJoin ){ 5445 /* For LEFT JOIN queries, cursor pIn->iCur may not have been 5446 ** opened yet. This occurs for WHERE clauses such as 5447 ** "a = ? AND b IN (...)", where the index is on (a, b). If 5448 ** the RHS of the (a=?) is NULL, then the "b IN (...)" may 5449 ** never have been coded, but the body of the loop run to 5450 ** return the null-row. So, if the cursor is not open yet, 5451 ** jump over the OP_Next or OP_Prev instruction about to 5452 ** be coded. */ 5453 sqlite3VdbeAddOp2(v, OP_IfNotOpen, pIn->iCur, 5454 sqlite3VdbeCurrentAddr(v) + 2 + bEarlyOut); 5455 VdbeCoverage(v); 5456 } 5457 if( bEarlyOut ){ 5458 sqlite3VdbeAddOp4Int(v, OP_IfNoHope, pLevel->iIdxCur, 5459 sqlite3VdbeCurrentAddr(v)+2, 5460 pIn->iBase, pIn->nPrefix); 5461 VdbeCoverage(v); 5462 } 5463 } 5464 sqlite3VdbeAddOp2(v, pIn->eEndLoopOp, pIn->iCur, pIn->addrInTop); 5465 VdbeCoverage(v); 5466 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Prev); 5467 VdbeCoverageIf(v, pIn->eEndLoopOp==OP_Next); 5468 } 5469 sqlite3VdbeJumpHere(v, pIn->addrInTop-1); 5470 } 5471 } 5472 sqlite3VdbeResolveLabel(v, pLevel->addrBrk); 5473 if( pLevel->addrSkip ){ 5474 sqlite3VdbeGoto(v, pLevel->addrSkip); 5475 VdbeComment((v, "next skip-scan on %s", pLoop->u.btree.pIndex->zName)); 5476 sqlite3VdbeJumpHere(v, pLevel->addrSkip); 5477 sqlite3VdbeJumpHere(v, pLevel->addrSkip-2); 5478 } 5479 #ifndef SQLITE_LIKE_DOESNT_MATCH_BLOBS 5480 if( pLevel->addrLikeRep ){ 5481 sqlite3VdbeAddOp2(v, OP_DecrJumpZero, (int)(pLevel->iLikeRepCntr>>1), 5482 pLevel->addrLikeRep); 5483 VdbeCoverage(v); 5484 } 5485 #endif 5486 if( pLevel->iLeftJoin ){ 5487 int ws = pLoop->wsFlags; 5488 addr = sqlite3VdbeAddOp1(v, OP_IfPos, pLevel->iLeftJoin); VdbeCoverage(v); 5489 assert( (ws & WHERE_IDX_ONLY)==0 || (ws & WHERE_INDEXED)!=0 ); 5490 if( (ws & WHERE_IDX_ONLY)==0 ){ 5491 assert( pLevel->iTabCur==pTabList->a[pLevel->iFrom].iCursor ); 5492 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iTabCur); 5493 } 5494 if( (ws & WHERE_INDEXED) 5495 || ((ws & WHERE_MULTI_OR) && pLevel->u.pCovidx) 5496 ){ 5497 sqlite3VdbeAddOp1(v, OP_NullRow, pLevel->iIdxCur); 5498 } 5499 if( pLevel->op==OP_Return ){ 5500 sqlite3VdbeAddOp2(v, OP_Gosub, pLevel->p1, pLevel->addrFirst); 5501 }else{ 5502 sqlite3VdbeGoto(v, pLevel->addrFirst); 5503 } 5504 sqlite3VdbeJumpHere(v, addr); 5505 } 5506 VdbeModuleComment((v, "End WHERE-loop%d: %s", i, 5507 pWInfo->pTabList->a[pLevel->iFrom].pTab->zName)); 5508 } 5509 5510 /* The "break" point is here, just past the end of the outer loop. 5511 ** Set it. 5512 */ 5513 sqlite3VdbeResolveLabel(v, pWInfo->iBreak); 5514 5515 assert( pWInfo->nLevel<=pTabList->nSrc ); 5516 for(i=0, pLevel=pWInfo->a; i<pWInfo->nLevel; i++, pLevel++){ 5517 int k, last; 5518 VdbeOp *pOp, *pLastOp; 5519 Index *pIdx = 0; 5520 SrcItem *pTabItem = &pTabList->a[pLevel->iFrom]; 5521 Table *pTab = pTabItem->pTab; 5522 assert( pTab!=0 ); 5523 pLoop = pLevel->pWLoop; 5524 5525 /* For a co-routine, change all OP_Column references to the table of 5526 ** the co-routine into OP_Copy of result contained in a register. 5527 ** OP_Rowid becomes OP_Null. 5528 */ 5529 if( pTabItem->fg.viaCoroutine ){ 5530 testcase( pParse->db->mallocFailed ); 5531 translateColumnToCopy(pParse, pLevel->addrBody, pLevel->iTabCur, 5532 pTabItem->regResult, 0); 5533 continue; 5534 } 5535 5536 #ifdef SQLITE_ENABLE_EARLY_CURSOR_CLOSE 5537 /* Close all of the cursors that were opened by sqlite3WhereBegin. 5538 ** Except, do not close cursors that will be reused by the OR optimization 5539 ** (WHERE_OR_SUBCLAUSE). And do not close the OP_OpenWrite cursors 5540 ** created for the ONEPASS optimization. 5541 */ 5542 if( (pTab->tabFlags & TF_Ephemeral)==0 5543 && pTab->pSelect==0 5544 && (pWInfo->wctrlFlags & WHERE_OR_SUBCLAUSE)==0 5545 ){ 5546 int ws = pLoop->wsFlags; 5547 if( pWInfo->eOnePass==ONEPASS_OFF && (ws & WHERE_IDX_ONLY)==0 ){ 5548 sqlite3VdbeAddOp1(v, OP_Close, pTabItem->iCursor); 5549 } 5550 if( (ws & WHERE_INDEXED)!=0 5551 && (ws & (WHERE_IPK|WHERE_AUTO_INDEX))==0 5552 && pLevel->iIdxCur!=pWInfo->aiCurOnePass[1] 5553 ){ 5554 sqlite3VdbeAddOp1(v, OP_Close, pLevel->iIdxCur); 5555 } 5556 } 5557 #endif 5558 5559 /* If this scan uses an index, make VDBE code substitutions to read data 5560 ** from the index instead of from the table where possible. In some cases 5561 ** this optimization prevents the table from ever being read, which can 5562 ** yield a significant performance boost. 5563 ** 5564 ** Calls to the code generator in between sqlite3WhereBegin and 5565 ** sqlite3WhereEnd will have created code that references the table 5566 ** directly. This loop scans all that code looking for opcodes 5567 ** that reference the table and converts them into opcodes that 5568 ** reference the index. 5569 */ 5570 if( pLoop->wsFlags & (WHERE_INDEXED|WHERE_IDX_ONLY) ){ 5571 pIdx = pLoop->u.btree.pIndex; 5572 }else if( pLoop->wsFlags & WHERE_MULTI_OR ){ 5573 pIdx = pLevel->u.pCovidx; 5574 } 5575 if( pIdx 5576 && !db->mallocFailed 5577 ){ 5578 if( pWInfo->eOnePass==ONEPASS_OFF || !HasRowid(pIdx->pTable) ){ 5579 last = iEnd; 5580 }else{ 5581 last = pWInfo->iEndWhere; 5582 } 5583 k = pLevel->addrBody + 1; 5584 #ifdef SQLITE_DEBUG 5585 if( db->flags & SQLITE_VdbeAddopTrace ){ 5586 printf("TRANSLATE opcodes in range %d..%d\n", k, last-1); 5587 } 5588 /* Proof that the "+1" on the k value above is safe */ 5589 pOp = sqlite3VdbeGetOp(v, k - 1); 5590 assert( pOp->opcode!=OP_Column || pOp->p1!=pLevel->iTabCur ); 5591 assert( pOp->opcode!=OP_Rowid || pOp->p1!=pLevel->iTabCur ); 5592 assert( pOp->opcode!=OP_IfNullRow || pOp->p1!=pLevel->iTabCur ); 5593 #endif 5594 pOp = sqlite3VdbeGetOp(v, k); 5595 pLastOp = pOp + (last - k); 5596 assert( pOp<=pLastOp ); 5597 do{ 5598 if( pOp->p1!=pLevel->iTabCur ){ 5599 /* no-op */ 5600 }else if( pOp->opcode==OP_Column 5601 #ifdef SQLITE_ENABLE_OFFSET_SQL_FUNC 5602 || pOp->opcode==OP_Offset 5603 #endif 5604 ){ 5605 int x = pOp->p2; 5606 assert( pIdx->pTable==pTab ); 5607 if( !HasRowid(pTab) ){ 5608 Index *pPk = sqlite3PrimaryKeyIndex(pTab); 5609 x = pPk->aiColumn[x]; 5610 assert( x>=0 ); 5611 }else{ 5612 testcase( x!=sqlite3StorageColumnToTable(pTab,x) ); 5613 x = sqlite3StorageColumnToTable(pTab,x); 5614 } 5615 x = sqlite3TableColumnToIndex(pIdx, x); 5616 if( x>=0 ){ 5617 pOp->p2 = x; 5618 pOp->p1 = pLevel->iIdxCur; 5619 OpcodeRewriteTrace(db, k, pOp); 5620 } 5621 assert( (pLoop->wsFlags & WHERE_IDX_ONLY)==0 || x>=0 5622 || pWInfo->eOnePass ); 5623 }else if( pOp->opcode==OP_Rowid ){ 5624 pOp->p1 = pLevel->iIdxCur; 5625 pOp->opcode = OP_IdxRowid; 5626 OpcodeRewriteTrace(db, k, pOp); 5627 }else if( pOp->opcode==OP_IfNullRow ){ 5628 pOp->p1 = pLevel->iIdxCur; 5629 OpcodeRewriteTrace(db, k, pOp); 5630 } 5631 #ifdef SQLITE_DEBUG 5632 k++; 5633 #endif 5634 }while( (++pOp)<pLastOp ); 5635 #ifdef SQLITE_DEBUG 5636 if( db->flags & SQLITE_VdbeAddopTrace ) printf("TRANSLATE complete\n"); 5637 #endif 5638 } 5639 } 5640 5641 /* Final cleanup 5642 */ 5643 if( pWInfo->pExprMods ) whereUndoExprMods(pWInfo); 5644 pParse->nQueryLoop = pWInfo->savedNQueryLoop; 5645 whereInfoFree(db, pWInfo); 5646 return; 5647 } 5648